Eve library

Eve — Deep Test & Impeccable Polish — TODOS (2026-08-04)

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Every aspect of Eve — every surface, state, tool, tour, voice path, string, and pixel — tested in deep detail, live in a real browser, in both themes, with every defect logged, fixed, and regression-locked. The initiative ends only when a full adversarial re-run finds ZERO defects (Phase 15).

Why this file exists: the 2026-08-04 live pass proved Eve's machinery (tools fire, cards render, rows land) but was honest about its blind spot — it never looked. The user then observed what the machinery pass stepped past: overlays burying chat messages, weird styling, weird copy. My own screenshots contain the evidence (seeded as defects in Phase 1). This file is the correction: function AND form, verified per item, no exceptions.

Process discipline (binding, per CLAUDE.md)#

  • One task at a time, sequentially. The checkbox is the sole source of truth. Each [ ] → [x] requires having performed THAT task's inspection in the current session. Never batch-mark. Never mark from memory of an earlier session.
  • Playwright is the visual harness (changed by the user, session 6 — supersedes "Claude in Chrome is mandatory"). The extension drives a real Chrome WINDOW, and on this Mac that window kept being occluded by the user's other windows: visibilityState goes hidden, requestAnimationFrame stops, and no transition, scroll or layout timing can be measured — the trap that produced the withdrawn EVE-VIS-014 and then blocked 2.2's narrow leg twice. It also cannot reach a 390px viewport, because a window that narrow is fully covered and cannot be raised programmatically (Chrome is granted at computer-use tier "read"). The harness lives at apps/oshun/web/e2e-inspect/ with its own config, outside e2e/ so nothing in it runs as a CI gate: npx playwright test --config e2e-inspect/playwright.config.ts <phase>. API-level or DOM-level verification still does NOT satisfy a visual checkbox — the harness screenshots every probe and attaches it as evidence. Claude in Chrome remains fine for exploratory looking; it is no longer what closes a checkbox.
  • A navigation screenshot is not an inspection. A visual task counts as done only when the screenshot was examined AT FULL ATTENTION for all five lenses — layout (alignment/spacing/overflow), overlap (z-index, obscured content), scroll behavior, contrast/legibility, copy — and the finding (or explicit "clean on all five lenses") is recorded. Use the zoom action on anything small (chips, timestamps, icons).
  • Both themes, always. Every visual task runs in light AND dark theme. A task verified in one theme is at most half done.
  • Three viewports for member surfaces: desktop (~1470w), narrow (~390w, iPhone-class), and desktop at 200% zoom. Admin surfaces: desktop plus narrow.
  • Console must be clean at the end of every page's inspection (read_console_messages onlyErrors:true) — hydration warnings count as defects.
  • Every defect gets a ledger entry (see the Defect Ledger section): id EVE-VIS-NNN, severity, surface, screenshot evidence, root cause, fix commit, regression lock (spec/e2e). A defect without a regression lock is not closed.
  • Fix-forward rule: severity-1 defects (content hidden, action impossible, data wrong) are fixed before proceeding to the next phase. Severity-2/3 may be ledgered and batch-fixed at the phase boundary — never deferred past it.
  • npx tsc --noEmit on the member web app is a real gate again (s5: "noCheck": true removed, 165 → 0). It needs NODE_OPTIONS=--max-old-space-size=8192. Run it before every commit that touches apps/oshun/web — three of this initiative's member-visible defects (borders dropped from a nonexistent token, a dead mood filter, a renderer that could not accept a real GPUDevice) were sitting in its output the whole time.
  • Commit + two-line push (branch AND main) at every phase boundary.

Harness recipe (from the 2026-08-04 run — battle-tested)#

  • The domain APIs are part of the stack, and this recipe omitted them until session 10. Tara and Arete are not in-process: the BFF proxies them to localhost:4101 and localhost:4401, and with nothing listening every Tara and Arete tool answers 502 domain_unavailable, so the agent reports ok: false and tells the member the room is "momentarily out of reach". That is why 3.3's battery recorded exactly that sentence as evidence of honest failure — it was honest, and its cause was environmental. Nyx and Nisaba need no service; they answer in-process. Boot the two that do:

    bash
    (cd apps/tara/api  && PORT=4101 TARA_DATABASE_URL=postgresql://oshun:oshun_dev@127.0.0.1:5432/tara   npx tsx src/index.ts)
    (cd apps/arete/api && PORT=4401 DATABASE_URL=postgresql://oshun:oshun_dev@127.0.0.1:5432/arete       npx tsx src/server.ts)
    

    Ports and the full six-service map: docs/audits/ANTIGRAVITY_TEST_PLAN.md §1.1. Verify with GET /v1/tara/recommended → 200, not 502.

  • The domain databases ship migrated and EMPTY. With the services up but no rows, every tool succeeds and shows nothing — which reads like a working happy path and is not one. node tools/seed-tara-content.mjs imports Tara's authored content tree (apps/tara/content) into the tara database: idempotent, nothing invented, and it skips-and-reports everything the tree names but does not contain (a collection's unwritten meditations, a course's unrecorded days). It yields 2 meditations, 1 course, 1 collection, 1 teacher — a small library, honestly. Arete has no equivalent importer yet, so its member lists are still empty.

  • Docker: postgres/redis/minio/mailpit up. BFF boot env: OSHUN_REDIS_URL, OSHUN_V1_DATABASE_URL, OSHUN_AUTONOMY_SNAPSHOT_KEY_BASE64 (+_REF), OSHUN_SIGNUP_VERIFICATION_HMAC_SECRET (≥32 bytes), OSHUN_AUTO_VERIFY_EMAIL=true, OSHUN_ASSISTANT_PROVIDER=openrouter, OSHUN_ASSISTANT_OPENROUTER_MODEL=deepseek/deepseek-v4-flash-0731, OPENROUTER_PROVIDER_SORT=price, OPENROUTER_API_KEY (from the user — never committed). Web on 3010 (CORS allowlist), admin on 3020.

  • Admin leg uses 127.0.0.1:3020 — localhost cookies are shared across ports and the member oshun-session cookie hijacks the signin route's token precedence.

  • Admin session: in-page fetch('/api/admin/signin', …) with x-oshun-customer-token: 'dev.'+btoa(JSON.stringify({sub: 'operator-studio-01', scopes:['admin:*','domain:*'], exp:…})) — build tokens IN-PAGE; the Chrome extension redacts literal token strings and blocks document.cookie.

  • Model cost rule (user directive, 2026-08-06; the id PINNED 2026-08-07). Every leg of this initiative, machinery AND conversation-quality, runs on the cheapest model that can do the job, through OpenRouter: OSHUN_ASSISTANT_PROVIDER=openrouter + OSHUN_ASSISTANT_OPENROUTER_MODEL=deepseek/deepseek-v4-flash-0731 + OPENROUTER_PROVIDER_SORT=price. Bind the dated snapshot, and know that it is not the cheapest — a session-13 claim that the alias cost 1.56× the snapshot was WRONG, and is corrected here: those were /v1/models summary fields, not routed prices. Priced properly against the live API (/endpoints per provider, usage: { include: true } for the billed rate), the alias is the CHEAPER of the two — $0.068/$0.137 against the snapshot's $0.102/$0.180 — because a slug on OpenRouter is an auction and the alias has cheaper third-party endpoints. The snapshot is bound anyway, deliberately (user decision 2026-08-08): this file closes defects on Fisher exact tests over arms run in different sessions — EVE-VIS-086 at 12/40 → 20/20 (p = 1.07e-07), EVE-VIS-094 at 29/40 → 11/40 (p = 1.0e-04) — and a model that can change underneath a floating alias makes those arms incomparable. The bigger lever is provider routing, which was not being sent at all: without it three consecutive calls went to three different providers at three different prices and quantizations. sort: price fixes both the price and the provider. Phases 0–4.3 ran on the alias with default routing; 4.4 onward runs on -0731, and every finding stays labelled with the id that produced it. The previous note here told a future session to sanity-check quality legs "with a stronger model when one is available" — priced out, that is $5/$25 per 1M against DeepSeek V4 Flash's $0.09/$0.18, i.e. ~56× input and ~139× output, or about $3.50 versus $0.06 for one 40-turn battery. Do not bind a frontier model to this harness. Findings stay labeled by model so a cheap model's own weirdness is never mistaken for a product defect; if a leg genuinely cannot be judged on the cheap model, ASK before spending rather than upgrading quietly. Keys: ~/Desktop/workspace/drawer/env-master.env (never committed). Full rule in CLAUDE.md § Model Cost for Tests and Harnesses.

  • Screenshots for evidence: save with save_to_disk and reference paths in the ledger. Tear down tracked PIDs at every session end.

  • The Chrome window must be VISIBLE, not merely open. Behind other windows the tab is visibilityState: "hidden" and requestAnimationFrame never fires: no transition completes, no programmatic scroll runs, and screenshots capture mid-transition geometry. Assert visible + rAF firing before ANY behavioural claim (snippet in the ledger's harness section). This produced one fully-worked false defect (EVE-VIS-014) before it was caught.

  • Pin the signup HMAC secret. The persisted signup-verification-state snapshot binds to a sha256 of OSHUN_SIGNUP_VERIFICATION_HMAC_SECRET; boot with a different string and the BFF dies with "Signup verification snapshot secret check does not match". This session's recovery: the snapshot held zero tokens/deliveries/challenges, so it was backed up and deleted, and the secret is now PINNED to oshun-eve-polish-signup-verification-hmac-secret-v1. Reuse that exact string.

  • Restart the web dev server after editing a workspace lib. Turbopack HMR does not always invalidate modules under libs/; an edit to libs/oshun/navigation surfaced as ReferenceError: … is not defined in the browser while the source on disk was correct.

  • Full member-web typecheck needs NODE_OPTIONS=--max-old-space-size=8192; the 4 GB default OOMs partway through the 3,186-file program.

Defect Ledger#

Lives at docs/audits/EVE_POLISH_DEFECT_LEDGER_2026-08.md (create in 0.4). Row format: | EVE-VIS-NNN | sev | surface | symptom | evidence | root cause | fix commit | regression lock | Severity: S1 = content hidden / action impossible / data wrong; S2 = clearly broken-looking or confusing; S3 = polish (spacing, tone, micro-alignment). ALL severities must reach "closed" for Phase 15 to pass.


Phase 0 — Harness, accounts, and evidence infrastructure#

  • 0.1 Boot the full stack per the harness recipe; verify the assistant binding live (one buffered API turn returns provider: openrouter). Record boot logs; the F2 workbench artifact sync line must be present and healthy. (2026-08-04: BFF+web up; sync line orphanFlagged: 0, verifier: []; live turn → openrouter/deepseek-v4-flash/agent. Landmine re-hit: signup HMAC secret must MATCH the persisted snapshot's secret across sessions — reuse dev-chrome-pass-…, don't mint fresh.)
  • 0.2 Create the standing test accounts: one fresh member (signup UI), one member with history (reuse eve-chrome-pass), one operator session (seeded operator-studio-01). Record ids in the ledger header. (2026-08-04: fresh = eve-polish-fresh@oshunsynthetics.com "Eve Polish Fresh" (onboarding intentionally parked at step 1/10 for task 12.1); history = eve-chrome-pass@oshunsynthetics.com (audit run 1/746 marked); operator = seeded operator-studio-01 via the proven signin recipe. Bonus finding: "USE ANOTHER ACCOUNT" clears the session with NO visual feedback until reload → EVE-VIS-009 (S3) ledgered.)
  • 0.3 Theme + viewport harness: confirm how the member app switches themes (settings/system) and script the three viewports; verify a screenshot in each theme×viewport renders distinguishably. (2026-08-04: three variants — cream (default light), dusk (dark), cream-hc — via data-lilith-theme attr + localStorage oshun-lilith-theme, system-resolved (prefers-contrast → cream-hc, prefers-dark → dusk). Switch in-page via localStorage.setItem + attr + oshun-lilith-theme-changed event. Both cream and dusk verified rendering distinctly on /welcome. Viewports: desktop 1470w; narrow = 500w EFFECTIVE (Chrome's window floor — noted, close enough to phone-class reflow); 200% zoom via document.documentElement.style.zoom approximation (Chrome blocks zoom shortcuts through the extension).)
  • 0.4 Create the Defect Ledger file with the seeded Phase-1 defects (EVE-VIS-001…008) and the evidence screenshots from the 2026-08-04 session (re-capture fresh ones — do not cite from memory). (2026-08-04: ledger live at docs/audits/EVE_POLISH_DEFECT_LEDGER_2026-08.md with 10 rows. VIS-001/002 re-reproduced FRESH in cream on the new account with saved evidence — containment measured (wheel over panel scrolled the page 0→500px; panel content unmoved; 1-turn greeting fully hidden behind follow-up cards). Two NEW rows: VIS-009 (use-another-account no feedback), VIS-010 (fresh account renders 7d-streak/42-sessions fixture stats as truth).)
  • 0.5 Investigate the Next.js dev-overlay badge ("3 Issues") on the member app: enumerate each underlying issue (hydration mismatch, console error, etc.), ledger each as its own defect with root cause. (EVE-VIS-007) (2026-08-04: overlay shows the issue class — React console error at design-system Button.tsx:325: mixed border shorthand + borderColor in one style object ("can lead to styling bugs" — React may drop the border color on rerender; a literal mechanism for intermittent weird styling). The earlier "3 Issues" count = the same error accumulating across pages this session; per-page sweep in 0.6. Root cause ledgered on VIS-007; fix scheduled with Phase 1.)
  • 0.6 Baseline console sweep: load /, /welcome, /domains/tara, /assistant, /assistant/audit, /assistant/graph and record every console error/warning per page before any fixing. This is the "before" picture Phase 15 is measured against. (2026-08-04: all six pages LOAD-time console-clean (fresh member session, cream). The VIS-007 Button error is INTERACTION-triggered (fires on rerenders during flows like signup), not load-time — Phase 15's certification must therefore sweep after interaction, not just after load.)

Phase 1 — Known defects from the 2026-08-04 live pass (fix FIRST)#

These are already-observed, evidence-backed defects. They block honest inspection of everything else (you cannot inspect a chat you cannot read).

Progress 2026-08-04 (session 2): 1.1 containment DONE and verified live (page 0px across wheel + full send; was 500px) via overscroll-behavior on body+messages, panel-only scrollTo (jsdom-guarded), sticky composer, hold + "Jump to latest" pill; 1.2 structure DONE (transcript before rail/meta; the 3,957px journey deck now a collapsed <details>; two e2e specs updated; 28/28 panel specs incl. the new hierarchy+collapse lock). OPEN: Playwright wheel-containment lock; the persistent-context STRIP deck above the transcript is still oversized — its spec-prose content converges with 1.3 (VIS-003), fix them together. New find: EVE-VIS-011 (Esc does not close the Keyboard Shortcuts modal). Neither 1.1 nor 1.2 is checkbox-complete until those close.

  • 1.1 EVE-VIS-001 (S1): panel transcript scroll. (CLOSED 2026-08-04 s3: containment verified 0px live twice + Playwright wheel lock green (assistant-panel-scroll-containment.spec.ts) + component lock; panel-only scrollTo, hold + jump pill, sticky composer.) The transcript does not auto-scroll to the newest message, and wheel-scrolling over the panel scrolls the PAGE behind it (broken scroll containment / no overscroll-behavior). Fix: newest message always visible on arrival (respecting a user-scrolled-up hold with a "jump to latest" affordance), panel scroll contained (overscroll-behavior, correct scroll container). Regression lock: component spec for auto-scroll on delta append + a Playwright check that wheel over the open panel does not move window.scrollY.
  • 1.2 EVE-VIS-002 (S1): follow-up cards bury the conversation. (CLOSED 2026-08-04 s3: fresh build verified — strip 428px compact, greeting visible on open, journey deck collapsed to a 45px disclosure row (body 7174px→3031px), conversation-first order locked by component spec; residual strip trimming → Phase 2.3/2.6.) The suggestion/continuity cards ("Return to Explore", "Preserve assistant excerpt…", "Reflect follow-up…") occupy the panel viewport above the transcript so replies are invisible without hunting. Decide and implement the hierarchy: conversation first-class, follow-ups collapsed/summarized (chips or a single expandable row) below the latest message, never between the user and an in-flight reply. Regression lock: spec asserting the latest assistant bubble is within the visible scroll area when follow-ups render.
  • 1.3 EVE-VIS-003 (S2): full-screen shell view — entry + chrome. (CLOSED 2026-08-05 s4: internal vocabulary purged from the member panel — Stage chip deleted, mode/avatar/fallback/transcript-state sections collapsed behind a "Session diagnostics" disclosure, "Persona switching…" rewritten in member voice. Live term scan over panel innerText: leaked:[] while the AI-assistant + session-memory disclosure stays visible; body 7174→1952px. Full-screen VIEW re-inspection rides Phase 2.6.) (a) Determine what click expanded the panel to the full-screen continuity view mid-conversation (my click at the input row triggered it) — if an expand affordance sits next to the input, review its hit target and affordance clarity. (b) The full-screen view exposes internal vocabulary to members: "Typed handoff", "Selection: 0 chars", "Mode state", "Streaming: Active", "Persona switching: Active". Rewrite or remove for member surfaces (builder/debug info belongs behind an operator flag). Copy per V1/BRAND.md voice.
  • 1.4 EVE-VIS-004 (S1, data): "20669d ago" date rot. (CLOSED-CORE 2026-08-04 s3: root = honest-unavailable briefing carries epoch-zero sentinels into formatRelativeTimestamp; formatter now refuses parsed<=0 and renders absolute month for >365d; 3-test lock green in HomeVeritasBriefingSection.test. The wider call-site sweep (7 more hand-rolled formatters found) folds into Phase 3.1 inventory.) Home evening catch-up card renders "⏱ 20669d ago" (epoch-zero into a relative formatter). Fix the source (unavailable timestamp must not format as 1970), and the formatter must refuse absurd inputs (>365d renders as a date or "—", never "20669d ago"). Then sweep EVERY relative-time call site in web+admin for the same class (grep the formatter's importers; test each with epoch-zero/undefined). Regression lock: formatter unit spec + the card's empty-state spec.
  • 1.5 EVE-VIS-005 (S2): tour dialog legibility. (MECHANISM CLOSED 2026-08-04 s3: dialog reads --l-surface-raised which was UNDEFINED — dusk fell back to a light card under light ink (light-on-light = the washed screenshots). Token now defined per theme (cream #fbf7ec / dusk #3a2f1f) + source-scan lock (lilith-theme-tokens.spec). Visual re-verify rides the Phase 5 live tour pass.) The TourPlayer dialog rendered semi-transparent over busy content — step text was near-illegible in two screenshots. Give the dialog an opaque (or heavily frosted) surface, minimum contrast per WCAG against any backdrop, and verify over the busiest page (Home). Check the spotlight scrim interaction (dialog must not be scrimmed with the page).
  • 1.6 EVE-VIS-006 (S2): admin Copilot drawer hierarchy. (2026-08-04 s3: transcript auto-scroll (container-only scrollTop on messages/actions change) + the entire meta wall (handoff/transcript/mode/persona/guards sections) collapsed behind a "Session details" disclosure; admin tsc + 9 specs green. Live drawer re-verify rides the Phase 9 admin pass.) Same transcript problems as 1.1/1.2 on the admin side (new messages below the fold; no auto-scroll) plus the meta-chip wall (TYPED HANDOFF / TRANSCRIPT STATE / MODE STATE blocks) pushing the conversation off-screen. Conversation gets the space; meta moves to a collapsed inspector. Regression lock: AdminAssistantChat spec for auto-scroll + a layout spec that the transcript container's height dominates the drawer.
  • 1.7 EVE-VIS-008 (S3): SW refresh toast overlap. (2026-08-04 s3: toast moved bottom-LEFT — bottom-right is owned by the assistant launcher and the /welcome signup card; component spec green.) The service-worker "Refresh to finish updating" toast overlapped the signup form's field area on /welcome. Reposition/auto-dismiss so it never covers interactive elements; verify at narrow viewport.
  • 1.8 Re-run the Phase 0.6 baseline pages after these fixes: capture the "after" screenshots, close the ledger rows with fix commits and regression locks. Commit + push both refs. (2026-08-05 s5: all six routes re-walked live in cream on a VISIBLE tab — /, /welcome, /domains/tara, /assistant, /assistant/audit, /assistant/graph — all 200, all load-time console-clean after a clean server restart. Ledger rows 001–013 moved to Closed with real fix-commit hashes and named locks. THREE NEW S1s found and fixed in this pass: VIS-012 (root host's duplicate launcher covering 72% of the dock's Send button — elementFromPoint returned "Open Lilith"), VIS-013 (the /domains "Preview surface" banner rendered UNDER the fixed sidebar and sticky top bar, so the one label saying "illustrative example data" was unreadable on the surfaces that render it), and the assistant half of VIS-016 (noCheck: true hid 165 type errors; 25 were on assistant surfaces, incl. 16 L.line — a token that does not exist — silently dropping every border on the graph explorer's scope gate). One finding WITHDRAWN: VIS-014 was a hidden-tab artifact (no rAF ⇒ no arrival scroll); two live negative controls disproved it and the speculative fix was reverted. New binding harness rule in the ledger: assert visibilityState === "visible" and rAF firing before trusting any behavioural observation.)

Phase 2 — Member panel: every state, pixel-inspected#

Every task: both themes × three viewports, console clean, five lenses.

  • 2.1 Closed state: the launcher/trigger on every shell variant (root AssistantHost mount AND ShellLayout-managed docks) — position, overlap with page content, focus ring, hover state, aria label. (2026-08-05 s5: BOTH variants inspected live, cream + dusk, at 1470w / 500w / 200% zoom. Variant A (ShellLayout trigger) — "Ask Lilith", 100×40 at (664,24) desktop, aria-label="Open AI assistant"; collapses to a 40×40 icon-only button in the compact bottom dock at 500w with the accessible name kept; at 200% it is 80×80, fully in viewport, clickable, no horizontal overflow; dusk repaints correctly (bg rgba(44,36,24,.4), ink rgb(192,173,141)) — note a first read showed cream values and was a REPAINT RACE, not a defect, so re-read after a frame. Focus ring present and identical to the rest of the shell chrome (two-tone box-shadow under :focus-visible; programmatic .focus() alone does NOT trigger it — press Tab first). Compact dock at 500w floats over scrollable content and covers no interactive target. Variant B (root host launcher) — 52×52 fixed at right:20/bottom:20, z-1300, aria-label="Open Lilith", meets 44×44, covers nothing interactive, dusk surface correct via --l-surface-raised. TWO DEFECTS FOUND AND FIXED: VIS-018 (no hover feedback on ANY of the six dock controls) and VIS-019 (no keyboard focus ring on the root launcher — the only way into the assistant on 55 shell-less pages). Also fixed here: VIS-004 reopened live on Home as "20670d ago" from a SECOND unguarded formatter; the guard now lives once in @/lib/time/relative-timestamp and every Home call site delegates to it.)
  • 2.2 Opening transitions: from launcher, from an invocation point, from a deep link (/assistant?prompt=…), from SelectionAsk — no layout shift of the underlying page, no flash of unstyled panel. (2026-08-05 s6: all four paths × cream/dusk × desktop 1470 / narrow 390 / 200%-zoom 735 — 6/6 green under the new Playwright harness (e2e-inspect/phase-2-2-opening-transitions.spec.ts), five lenses each, console clean (zero page errors, zero app errors, zero failed requests), screenshot attached per probe. SIX defects found and fixed: VIS-022 (S1) the transcript was a nested scroller so a real thread was NOT ON SCREEN at all (0 of 8 hit-test points hit a message; the panel showed persona cards); VIS-026 (S1) at 390px the SW update toast covered the only launcher — elementFromPoint returned its "Later" button — a regression from VIS-008's own fix; VIS-020 (S2) the page re-flowed in one frame onto a 265px empty gutter while the panel still had 500ms of entrance to run (max gutter now 1px, largest shift 0.1513→0.0716); VIS-021 (S2) ?prompt= was ignored and a different question was sent as the member's own; VIS-025 (S2) the Send glyph at 2.53:1 and the member's own bubble at 1.73:1 in dusk; VIS-023 (S2) the selection chip said "Ask Oshun". Every lock calibrated by mutation. Two findings were the HARNESS, not the app, and were withdrawn before they reached the ledger: a hydration mismatch caused by planting the theme attribute pre-hydration, and overlap reports for the launcher sitting under an aria-modal panel and under Next's dev-only portal. VIS-027 (dusk contrast on the Tara room's own chrome) was found here but belongs to Phase 12.3/12.7 — verified pre-existing with the assistant never opened, and ledgered rather than fixed in place.)
  • 2.3 Docked panel — empty state: greeting copy, suggestion buttons, disclosure chips (Memory/Grounding/Persona/Disclosure) — chip copy reviewed word-by-word for member-appropriateness; spacing/wrap at narrow width. (2026-08-05 s6: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-2-3-empty-state.spec.ts) — nothing clipped, nothing spilling the panel, no control covered, all text ≥ WCAG AA, console clean. Method: dump EVERY rendered string, split into member-visible vs collapsed/sr-only (the ~4,000px journey deck is legitimately closed per 1.2 — counting it would have produced ~165 false findings), then read the visible list word by word. 60 visible strings, 18 of them ours. VIS-015 CLOSED and two new rows: VIS-028 (S1) a fresh session asserted "Applied handoff to Support Assistant under Session memory boundary" — an action the member never took, because the renderer invented an audit event when the real list was empty; VIS-029 (S2) the switcher said "Lilith is guiding this conversation" while the block beneath said "Support Assistant" (two resolvers, two defaults) — the member is now told one thing, and the id-space divergence is ledgered for the Phase 2 boundary. Also rewritten: the raw invocation id ("Customer Web.Global Launcher") that appeared twice, the Anchor/Thread/ Entry/Evidence/Tools/Artifact chips (moved into the Session diagnostics disclosure), the "Context handoff" card (collapsed — e2e specs assert on its chips), the four-pill wall on every bubble including the member's own, "Shell navigator", "Active persona", and "Current thread only". 42 visible strings now. Residual S3 noted for 13.3: "Session memory" and its sentence appear twice, and "You're in Lilith." reads oddly on Home.)
  • 2.4 Docked panel — conversation state: user bubble, streaming bubble (delta growth without jitter), completed bubble, tool-note line ("Consulted: …"), timestamps, long-message wrapping, code/list markdown rendering, link styling, copy-to-clipboard if present. (2026-08-05 s7: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-2-4-conversation-state.spec.ts), five lenses, console clean, screenshot per probe. Method: a scripted turn streams over a real ReadableStream at the provider boundary — fulfil delivers the whole body in one tick, which would make "delta growth" unmeasurable — while the transcript is recorded EVERY animation frame (chars, top, opacity, scrollTop, bubble identity, typing-dot presence, document scrollWidth). NINE defects, all fixed and locked. The first blocked the task outright: VIS-030 (S1) no member could hold a conversation at all since yesterday's four-room cut — availableDomains was an ordering hint rather than a restriction, so the BFF wrote deferred rooms into every session and the release-scoped store refused it; every message answered "I'm having trouble reaching the assistant service right now". Then, in the conversation itself: VIS-031 (S1) a stored grounding level this build does not know took the whole page into the error boundary (exhaustive switch, no default, behind a sanitize that was a cast); VIS-034 (S1) every reply rendered as one <p> — newlines collapsed, ```and- and[label](url)printed out, no clickable links, and a 130-char URL painted 216px past the panel's right edge; VIS-035 (S1) the panel read its own follow-the-reply scroll as the member scrolling up and stranded a growing answer 827px below the fold; VIS-032 (S2) the finished reply blinked out and jumped 27px because completion minted a new React key; VIS-033 (S2) typing dots ran under two paragraphs already written; VIS-036 (S2) nothing told the member the answer came from their own saved reading; VIS-038 (S2) dusk timestamps at 2.94:1; VIS-037 (S3) the transcript ghosted through the 97%-opaque composer. Replies now render through a closed-subset renderer (lists, fenced code, links, emphasis) that can never produce markup and refuses any scheme but http/https/mailto, plus a Copy control that says "Couldn't copy" rather than faking success. Its own spec caught two bugs in it: a shared/g/regex used recursively looped until the heap died, and\1inside the bold branch pointed at the code-span group, so **important**rendered as a bold "i". Every lock calibrated by mutation. THREE findings were the harness's own fault and never reached the ledger — a recorder locked onto the typing-dots bubble (reporting the reply as never growing), an analyser comparing the greeting's numbers against the reply's, and the covered-control lens applied to a content block taller than the viewport at 200% zoom. Noted for later: 117 more #6f6555literals remain inglobals.css (→ 12.7), and 4 BFF home-route tests were already red from the release-cut commit before this session.)
  • 2.5 Docked panel — input row: placeholder copy, disabled/sending states, mic button states, send button affordance, focus trap behavior, Enter vs Shift+Enter, long input growth, character handling (paste a 2k-char text; paste an emoji-heavy text). (2026-08-05 s8: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-2-5-input-row.spec.ts), five lenses, console clean, screenshot per probe. Method: every claim is a keystroke followed by a measurement — a real clipboard paste (method recorded, insertText fallback), a scripted slow stream so the in-flight window can be sampled six times, and an Enter hammered mid-flight to see whether the turn duplicates. EIGHT defects, all fixed and locked. The composer was an <input type="text">, which is three findings in one: VIS-039 (S1) 848px of a 1084px question scrolled off to the left where the member could not read their own words, and a pasted quotation lost all 7 line breaks — the message sent was not the message pasted; VIS-040 (S2) Shift+Enter was a dead key; VIS-044 (S2) the box carried no placeholder at all, its only prompt a 1px screen-reader hint. It is now a textarea that grows to six rows and then scrolls. Then VIS-041 (S1): the app blocked its own microphone — Permissions-Policy: microphone=() on every route, so push-to-talk, server STT and Web Speech all died with "Permissions policy violation" before the member was ever asked, while the mic button flipped to "Stop listening" over a greyed-out composer; the whole of Phase 6 was unreachable. VIS-042 (S2) every voice failure was silent (three bare returns and an error handler that only flipped state, plus no bound on a recognizer that answers with nothing). VIS-045 (S1) at 390px the panel claims aria-modal="true" and 14 of 14 Tabs left it for invisible page content. And from the evidence screenshots themselves: VIS-046 (S2) an uppercase SHELL badge in the header on every non-room route (DOMAIN_LABELS fallback), VIS-047 (S3) "Return to Home" and "Back to Home" side by side pointing at the same href. Every lock calibrated by mutation. THREE findings were the harness's own and never reached the ledger: a send button read mid-transition (0.3 opacity was the animation's start value), a 160-char "long" question that fits on one line in an 1157px composer, and Next's dev-only <nextjs-portal> intercepting clicks at 390px. One of my own fixes had a real bug the browser caught — sizing a border-box field from scrollHeight clipped 2px of every row. Also fixed here: e2e/assistant-transcript-in-view.spec.ts was RED before this session (its double routed /message while the panel streams /turns, and fulfilled cross-origin responses carried no CORS headers, so the double had never once been used). Ledgered not fixed: VIS-043 (font preload hints — measured, typography is correct, re-check on a prod build at 15.3), VIS-048 ("Shared shell"/"Shell surface" inside disclosures → 13.2), VIS-049 (Home fires a dead Veritas route 3× and a failing Arete route 10× per load → 12.2).)
  • 2.6 Full-screen shell view (post-1.3): every card, every string, both entry and EXIT paths (Esc, close button, back), state preserved on exit. (2026-08-06 s8: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-2-6-fullscreen-view.spec.ts), five lenses, console clean. First the surface had to be located honestly: at 1470 a shell page gives the member the DOCK (it mounts at ≥1360px), so the overlay presentation is reachable at desktop ONLY from a shell-less route via the root host launcher — /legal/terms — while at 390 and 735 the shell's own trigger opens it, which is what every phone member gets from every page. Each cell records which entry it used. THREE defects, all fixed and locked. VIS-050 (S2): the overlay covered the entire desktop viewport — 1,470px wide instead of its own 420px sheet, giving ~1,400px lines of body copy — because an inline width: '100%' added for the dock applied to both presentations and had made the stylesheet's sheet width dead code. VIS-052 (S2): every exit (Escape, close button, scrim) dropped focus on <body>, so a keyboard member who dismissed the assistant had the whole page to tab through to reach the launcher again; the fix needed two goes — a synchronous restore runs before the root host re-renders its launcher, and the captured "opener" was body itself on the desktop path, which focuses to nothing. VIS-051 (S3): the header said Lilith · LILITH · LILITH. Verified working, not just asserted: all four exits close, the page is never left scrimmed or unscrollable after Back, and the thread (1 user + 2 assistant turns) survives close→reopen intact at every viewport. Two harness lessons folded in: the inspection config now sets actionTimeout (a stuck click had been eating the whole 240s test budget — 17 minutes for one cell), and the scrim leg skips itself below 768px where the panel IS the screen and there is no scrim to click. Noted for 13.3: on a shell-less route the context card still says "You're in Lilith." and offers "Return to Lilith".)
  • 2.7 Error / degraded states rendered honestly and beautifully: BFF down (offline banner + preserved transcript), 503 agent_not_configured (deterministic fallback — verify the member sees no scary jargon), budget exhausted (429 copy), refusal (turn.error fallback path), voice-not-in-plan 403 copy. (2026-08-06 s8: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-2-7-degraded-states.spec.ts) — all five faults driven in one pass per cell with the BFF's OWN payloads, and judged on both sides: nothing fabricated, and nothing leaked (an 18-term sweep for status codes, reason slugs, provider names, undefined). Results: the 503 agent_not_configured, the 429 budget refusal and the turn.error refusal all hand the member a real deterministic answer with no trace of the failure, and no grounding badge is ever minted for a turn that did not happen. TWO defects, both fixed and locked. VIS-053 (S1): going offline made the assistant impossible to type into — the PWA offline card sits bottom-right at z-60 and landed on the dock's composer at z-11, so elementFromPoint over the input returned the card's heading; bottom-right belongs to the assistant, and every other floating status surface had already been moved to bottom-left (VIS-008/026) while this one was missed. VIS-054 (S2): a plan-gated voice 403 was reported as "That recording could not be transcribed… Please try again" — the recording was fine, the plan does not include server transcription, and the member was sent back to a control that will refuse them every time. TWO of my own findings were withdrawn before they reached the ledger, both because the double was unfaithful: a 429 on /message "proved" that an exhausted budget tells the member to retry in a moment, but refuseBudget exists only on the turns route and the deterministic engine answers instead; and "no offline banner" was measured by aborting requests, which leaves navigator.onLine true — context.setOffline shows the banner correctly. The harness now runs Chromium with a fake media device, without which every server-voice probe silently measures a recording that never happened. Noted for 3.3: under the 503 fallback the deterministic engine answered a Tara question with tonight's sky.)
  • 2.8 The panel over every page CLASS it can open on: Home, a domain room (each of the 6), a content/reader page, settings/profile, /assistant/audit, /assistant/graph — check the panel never obscures primary page actions unrecoverably and page scroll behind is correctly locked/allowed per design. (2026-08-06 s8: 5/5 green (e2e-inspect/phase-2-8-page-classes.spec.ts) — nine page classes × cream/dusk × desktop/narrow/zoom200, 45 page-opens in all. Rooms are the four the V1.0 release scope keeps (Tara, Nyx, Arete, Nisaba); walking six would have spent a third of the run on 404s. Method: for each page, enumerate ITS OWN in-viewport controls, hit-test every one of them, and judge by presentation — the non-modal dock must cover nothing (measured: 0 covered on all nine classes at desktop and zoom200), while the phone-width overlay covers the page by design and is required to give it back, which Escape does on every class. Page scroll behind the modal never moved; no class scrolls sideways; the panel's own body scrolls everywhere. One defect, fixed and locked — on a page the assistant merely opens over, not on the assistant itself. VIS-055 (S2): /settings threw an uncaught React hydration failure, the server sending Jul 21, 2026 – Aug 20, 2026 for the latest invoice while the client rendered Jul 22, 2026 – Aug 21, 2026; the default billing snapshot derives its dates from new Date() at call time and is rendered on both sides of hydration. Anchored to the start of the UTC day. Two harness corrections: the first version measured window.scrollY for "does the page behind scroll", which several shell routes would have answered falsely because they scroll an inner container; and the matrix run that "proved" the hydration fix ran at 06:13, hours from the boundary that causes it — so the real lock is a fake-timer unit test in libs/oshun/auth, not the probe that found it.)
  • 2.9 Accessibility pass on the panel: focus order, aria-live on streaming replies, screen-reader labels on all controls, reduced motion honored, 200% zoom usable. (2026-08-06 s8: 6/6 green (e2e-inspect/phase-2-9-accessibility.spec.ts) across cream/dusk × 1470/390/735, each property measured rather than eyeballed. Clean: all 31 operable controls in the panel carry an accessible name; the forward tab walk from the panel's first focusable stays inside it for 16 stops with no invisible stop and no dead end; reduced motion collapses every animation to 0.01ms (panel entrance, message entrance, transitions); axe reports 0 violations at every size. One defect, S1, fixed and locked: VIS-056 — the transcript had an aria-label and a tab stop but no live region anywhere in the panel, so a screen-reader member sent a question and heard nothing; the reply landed in silence and had to be hunted for. The typing indicator was the same thing in miniature: three animated dots with a null accessible name. Worth recording that axe passed both before and after — a missing live region is not a rule it checks, so this was only findable by asking the question directly. THREE of my own findings were withdrawn before the ledger, all measurement faults: "nine nameless controls" (Chrome returns innerText: '' for off-screen content-visibility subtrees — they were simply scrolled out of the panel body, and textContent shows their labels), "the panel animates under reduced motion" (test.use({ reducedMotion }) never reached matchMedia; page.emulateMedia does), and a tab order that appeared to leave immediately (the walk started at the composer, which is LAST in the panel's DOM). Noted for 12.6: one focusable a[Home] on the shell measures 0×0 while focused.)

Phase 3 — Conversation copy & content quality#

  • 3.1 String inventory: extract every member-facing string the assistant surfaces can render (panel components, system notices, chips, tooltips, empty/error states, tour dialog, voice states, confirm cards, board pages). Store the inventory with file:line references beside the ledger. (2026-08-06 s9: 2,094 distinct strings across 62 files and 2,338 call sites, mined by tools/build-assistant-string-inventory.mjs into docs/audits/EVE_POLISH_STRING_INVENTORY_2026-08.{json,md} beside the ledger. Built as a re-runnable tool, not a one-off dump, because 13.1 has to diff it: extraction is AST-based (JSX text, attributes, object properties, variables, call arguments, and TEMPLATES rendered with {…} for their spans, so You're {…}% through "{…}" — {…} of {…} lessons done. is one row rather than four fragments), verdicts live in a separate sidecar keyed by a hash of (surface, text) so regenerating never destroys editorial work while REWORDING a string retires its verdict, and --check fails on staleness. The filter was calibrated in both directions, and both directions found errors. Against false negatives: a raw quoted-string scan of every file, differenced against the inventory, with every survivor read in context — that caught a structural-name list which was silently eating const prefix = isMorning ? "Let's start your morning routine!" : …, and the list is now four times shorter because ids and slugs are rejected by their SHAPE, which does not depend on guessing property names. Against false positives: reading the kept rows for the small surfaces where noise cannot hide, which produced three precise rules (anything inside a style={{…}} object is CSS whatever its property is called; 'use client'; KeyboardEvent.key comparands) rather than a heuristic that could also swallow "Step 2 of 6". Every rejection is counted by rule in the JSON header and dumpable with --rejects — a filter nobody can audit is a filter nobody should trust. Scope decisions recorded in the tool rather than left implicit: domain-intents.ts is OUT (491 strings of NLU matcher input — examples/keywords/an internal description that no render path reads; inventorying "I keep waking up at night" would put 491 rows in front of an editor no member will ever see), the BFF's agent-tools.ts is IN (the do-tier confirm card's sentence is authored server-side, so a web-only inventory would miss every one of them), and admin-agent-tools.ts is out as builder tier. Two things the inventory surfaced on its own, carried forward rather than fixed here: only 4 mutating member tools exist against the 7 that task 4.5 names (no remove-favorite, unsave, unfollow or set-reminder binding) → 4.5, and a mutating tool with no confirmSummary shows the member a card reading "Run nyx_log_observation." → 3.4.)
  • 3.2 Editorial pass over the inventory per V1/BRAND.md: Lilith voice for members (calm, no engineering jargon, no internal vocabulary), Eve only ever builder-side; flag+fix every violation (the Phase 1.3 finds are the known seed). (2026-08-06 s9: seven defects, EVE-VIS-057…063, all fixed and mutation-locked, in four commits. The largest was not a string but an identity: the system prompt introduced Lilith as the keeper of six rooms and named the two V1.0 does not open, in three places the release cut had never reached — the identity line, every daypart's "users often want to" list, and the personality guideline — while only the capabilities section and the tool definitions were scoped. The model was therefore handed a worked example of naming a room that 404s ("Veritas pulled the sources"), and told four times a day that this member often wants to check the news on Veritas. Beside it the deterministic engine said "I keep six rooms here" directly above the four bullets it had already scoped; navigate advertised all six as an enum and refused afterwards, which is a refusal arriving after the promise; and a curated tour offered "check claims, find courses". The second wave was the instructive one. Re-running the miner after the first fix and sweeping again found three more replies — the greeting, the catch-all, and the domain-switch answer — and the catch-all's violation had no room name in it to grep for: "You can ask me about meditation, news, the night sky, or your goals" is Veritas by capability. That is the reason the fix was not "delete two names" anywhere: every one of these now assembles from authorizedDomains, so V1.2 restores them by opening the rooms. Separately: audience was declared on all four curated tours and read by nobody, so list_curated_tours offered every member "Build with the assistant" and its workbench-queue narration — a field that looked like a gate and gated nothing; the guide switcher's chips advertised deferred rooms on every card including Lilith's own ("Shell · Library · Nyx · Arete · Metis") in a section that is not behind a disclosure; and the continuity rail explained our architecture to the member in our vocabulary. The panel spec that should have caught that last one stripped closed <details> before checking — which is why it passed while three occurrences sat inside them (the deferred EVE-VIS-048). Lifting the exclusion found five MORE, in a different package. A disclosure is collapsed, not private. Scope stated rather than quietly widened: "All domains" also appears in Library, Explore, Search and Notifications — not assistant surfaces, so they stay with 13.2. What was READ, precisely, since the inventory is 2,138 strings: every string on the member panel and on all 20 small surfaces, one by one; every string on every surface against seven rules (Eve, OSHUN, deferred room, six-rooms, domain, shell/handoff, capability-without-a-name), with each hit opened in its source and its reachability decided rather than assumed — that is how the engine replies Veritas strings were correctly left alone (keyed by a domain a member's session does not carry) and the greeting's were not. The final sweep returns only model-facing JSON-schema text and strings inside the Session diagnostics disclosure. Not done here, and it is 13.1's task, not a silent gap: the per-string verdict sidecar is still empty. Recording 2,138 verdicts belongs after all fixes land, which is exactly what 13.1 says; writing them now for strings I judged by rule rather than by reading would be a worse artifact than an honest zero.)
  • 3.3 Live conversation quality battery (the model IS part of the surface): 10 scripted member conversations across the 6 domains + shell (greeting → domain question → follow-up → tool-grounded answer). Judge each transcript for: grounded claims only, cited sources when docs were searched, honest "I can't" when tools fail, no invented UI references, tone. Ledger every weird reply with the prompt that produced it. (2026-08-06 s9: 10 conversations, 27 turns, zero errors, all agent-mode on a live model, driven at the BFF's real streaming turn route and judged from the transcripts — committed at docs/audits/evidence/EVE_POLISH_3_3_BATTERY_2026-08-06.json, script at tools/eve-conversation-battery.mjs so it re-runs at 15.1. Model: deepseek/deepseek-v4-flash per the user's cost rule; every finding below is model-attributed. Four rooms, not six — the session's own authorizedDomains comes back [tara, nyx, arete, nisaba], so scripting Veritas and Metis conversations would have spent a fifth of the battery proving that 404s 404; the two rooms are instead probed adversarially by a conversation that tries to bait them out. API-level rather than Playwright, deliberately: 3.3 judges what the model SAYS and Phase 2 already covered how the panel renders it — the "API-level does not satisfy a checkbox" rule is scoped to visual tasks. What passed, and it is the more important half: across 27 turns and 15 distinct tools, not one model reply named Veritas or Metis — including the conversation built to pull them out ("What's in the news today?" → "I don't have a news domain, I'm afraid — my house is about meditation, the night sky, daily growth, and contemplative reading"; "What courses do you have?" → Tara and Arete only). That is EVE-VIS-057/062 verified against a live model rather than a unit test. Honest failure was strong throughout: tool outages reported as outages ("both Tara and Arete are momentarily out of reach"), absent history reported as absent and then backfilled with what it COULD see (two real Nisaba workspaces, an empty Nyx log), out-of-scope asks refused cleanly (weather, email). Three defects, EVE-VIS-064…066. The serious one is VIS-064 (S1): a member's grounded doc answers come from the engineering docs centre — the member corpus is docs-center/** on the assumption it is curated, but its 6,486 entries are 3,450 package/library docs, 1,348 ComfyUI workflows, 345 API-reference pages and 7 tagged V1, so Lilith told a member "Analytics and call-sheet exports exist at the API level" and "billing … handled through Stripe" (a vendor, and a rail V1.0 does not ship). That row stays OPEN: which of 6,486 docs a member may search is a content-scope decision, not a code fix, and 4.9 owns the corpus boundary — raised for the user rather than guessed at. Shipped now is the half that does not depend on it, plus VIS-065 (S2): the model described screens nobody verified exist, always hedged ("usually in the top corner", "your profile or stats page") and therefore easy to miss — guideline 6 forbade inventing DATA and said nothing about the interface. Two new prompt guidelines close both, mutation-locked. VIS-066 (S3): all ten sessions carried continuity.grounding.summary = "Assistant session ready for grounded Metis study handoff." — a room-specific module serving as the generic builder; checked rather than assumed to be unrendered, then fixed anyway.) RE-RUN 2026-08-06 s10 (user-approved), and it changed the verdict. The original battery was judged against a stack where the Tara and Arete domain services were not running at all — a gap in this file's own harness recipe, now fixed above. Half of every tool call in it had failed: 22 of 44. Re-run with the services booted and Tara's content imported: 5 of 40, and all five survivors were one tool, which turned out to be a real defect (EVE-VIS-072). What the original pass got RIGHT is unaffected and is now confirmed against an environment where the rooms actually answer: still not one mention of Veritas or Metis across 27 turns, so EVE-VIS-057/062 hold. What it could not have seen is the whole happy-path half — the re-run's replies name real sessions ("Calm Your Mind — 10 minutes, low intensity"), the real daily passage ("Attention Before Interpretation"), and stay honest about absence ("Doesn't look like you've favorited or have an in-progress session saved, so I can't say for certain"). Two S1s came out of reading those transcripts, both fixed and locked: EVE-VIS-072 (arete_active_goals had never worked for anyone — a type that never asked for the member) and EVE-VIS-073 (a member who asked what was in the sky tonight was told the moon was new and the sky perfectly dark, when it was last quarter at 45% — the "tonight" tool returned a fortnight of future events and nothing about tonight). Neither was findable before, because both were hidden inside the same symptom as the dead services: a tool reporting failure, or a room reported as out of reach. Re-run artifact: /tmp/battery-3-3-rerun.json; the committed 2026-08-06 evidence file is left as the original record. Still open from the first pass: EVE-VIS-064 (the docs a member can search are the engineering docs) — a content-scope decision that 4.9 owns, unchanged by any of this._
  • 3.4 Suggested actions / follow-up card copy: every card variant that can render, reviewed for usefulness and voice; verify each card's action actually does what its label says. (2026-08-06 s9: the second half of that sentence found an S1. EVE-VIS-067: the BFF and the panel hold two different definitions of a card — the engine's AssistantResponseCard over seventeen type variants against the panel's ResponseCard over six, sharing exactly two union members and, of the fields the panel renders, only title. Nothing mapped between them: the route forwards cards: response.cards verbatim and the panel assigned data.response.cards ?? [] straight across, a cast tsc cannot see because the payload is JSON. So an engine card rendered as <a class="assistantCard"><strong>Evening Wind-Down</strong><small></small><em> ›</em></a> — subtitle dropped, no href, and a chevron whose handler ran window.location.href = card.actionHref, which coerces undefined to the string: the member's panel closed and dropped them on /undefined. Cards come from the deterministic engine, which 2.7 already proved members reach. The instrument was wrong before the app was: my first lock asserted the literal string "undefined" would appear and PASSED against the broken build, because React renders undefined as empty — the real shape only came out by forcing the DOM into an assertion message. Fixed by normalising at the boundary (both ingestion points, including transcript restore) and by making a card with no destination a div with no chevron rather than an anchor promising one. Also EVE-VIS-068 (S2): card titles are pass-throughs whose fallbacks are the copy, and the achievement card's title was u.achievementId — a machine id as a heading, the EVE-VIS-015 class again; a nameless Nyx hit titled itself "Object". Both locks mutation-calibrated, and the second carries a vacuity guard that fired on its first run and caught the test reaching no card at all. Two harness lessons paid for: the panel restores its transcript from sessionStorage, so a second test read the first test's card until it cleared the right store; and the live battery is what established that the agent path emits zero cards in 27 turns, which is why this defect had never been seen. Left as recorded S3s rather than widened into here: a workspace card subtitled with its raw lowercase kind, and the Metis catalog card's literal 'Metis' subtitle fallback (unreachable — the room's tools are not in a V1.0 member's toolset).)
  • 3.5 Formatting torture tests rendered correctly: model replies containing tables, nested lists, long code blocks, RTL text, very long unbroken strings (URLs) — no horizontal blowout of the panel (this is a known failure class in chat UIs). (2026-08-06 s10: 6/6 green across cream/dusk × 1470/390/735 (e2e-inspect/phase-3-5-formatting-torture.spec.ts) — five torture replies driven through the real composer per cell, 30 probes, five lenses each, console clean, a screenshot and a measurement JSON attached per probe. Two of the five cases were confirmed reachable against the live model before a line of the spec was written, because the alternative was fixing a rendering path no reply can take: asked for a table, deepseek/deepseek-v4-flash returns four pipe rows and a |------| delimiter row, and a member who writes Arabic is answered in Arabic — 263 RTL characters with an ordered list and a mixed Arabic/Latin line. Both torture replies below are those live replies verbatim. The headline invariant the task names — no horizontal blowout — was already held, and that is the more important half: at every level measured (page, panel, transcript, bubble, and every descendant against the bubble's content box) a 268-character URL and a 300-character unbreakable word stayed inside, bare, as a link label, in a list item, in a heading, in a quote, in inline code, in a fenced block, and in the MEMBER'S OWN bubble, which renders through a different path and had never been asked. 2.4's overflow-wrap: anywhere covers all of it. The 40-line code block was clean too. Three defects, all fixed and mutation-locked. EVE-VIS-069 (S2): the block parser read no indentation at all, so a list inside a list was destroyed two different ways by the one cause — under a numbered parent the sub-steps were glued into the parent's own line with their - markers showing (1. Monday - Ten minutes of breath work - One passage), and under a bulleted parent they were PROMOTED to siblings, so "Miss one day, not two" stopped being a detail of "Consistency over length" and became a peer of it. That is meaning changed, not styling. EVE-VIS-070 (S2): no table branch existed, so the model's table arrived as prose and wrapped mid-row in a 293px bubble — pipes through the middle of sentences, columns gone, a line of bare dashes where a rule should be. The renderer's own comment had reasoned that source text was "honest rather than half-drawn", which is true of a half-drawn table and not of |------|. EVE-VIS-071 (S2): nothing in the transcript carried dir, so Arabic inherited the document's ltr — measured by comparing the client rects of a sentence's last two characters, the full stop rendered at the HEAD of the sentence instead of its end, and an Arabic list kept its numbers and indent on the left, detached from the items. My own fixes had two bugs each caught by a real check rather than by review. The unit spec caught the list parser hanging a kind-change list inside an empty <li> of the list before it (four items counted in a two-item list) and a delimiter pattern requiring two dashes, which silently rejected |:-:|. Then the BROWSER caught what no unit test could: dir="auto" on both the list and its items is worse than redundant — resolution SKIPS descendants carrying their own dir, so the items blinded the list, the <ol> fell back to ltr, and its markers spilled 17px out of the list box; per-item direction belongs to unicode-bidi: plaintext, and the indents had to become logical (padding-inline-start, border-inline-start). And the first table I drew was measurably unreadable — a header rendering "LAST TOU CHE D" and a cell "Compare translatio n workspac e" — from overflow-wrap: anywhere collapsing the column min-widths, plus the design system's base th (11px, letter-spaced, uppercase) turning the reply's own "Name" into "NAME". Both are now measured every run, not eyeballed: the harness reports the header text as rendered and every word broken across lines inside a column wide enough to hold it. Regression locks are unit specs in CI, not this harness: assistant-markdown.spec.tsx 19→44 cases (two of the table cases are NEGATIVE controls — a pipe in prose and a fenced ASCII diagram must NOT become tables) plus one in AssistantPanel.test.tsx for the member's own bubble; each of the four new groups was calibrated by mutating the source and watching it go red. npx tsc --noEmit clean, 156/156 assistant component specs green.)

Phase 4 — Every member tool, live and visually verified#

For EACH tool: trigger it through a natural conversation, verify the visible result (not just the API), inspect any UI it drives, and verify the honest-failure rendering (break the dependency where feasible).

  • 4.1 Domain read tools — tara (recommended sessions, continue, progress, favorites list, session audio, health). (2026-08-07 s11: 24 probes, 6/6 theme×viewport cells, live model and live tools (e2e-inspect/phase-4-1-tara-read-tools.spec.ts), plus a second spec run with the Tara service actually stopped. Two of the six things the task names cannot be reached at all: getSessionAudio and getHealth are declared on the assistant's Tara adapter and wired in the route, and then nothing calls them — no binding in agent-tools.ts, no case in the deterministic action-router (EVE-VIS-074, S3). The other four were driven through the real composer as natural questions. Two things had to exist before any of this meant anything. The domain databases ship empty, so continue/progress/favorites had only ever been seen saying "nothing yet" — tools/seed-tara-member-history.mjs gives a member a favourite, a part-finished sit and a course enrolment, through Tara's own facade for the write it has an endpoint for. And ground truth is READ OUT OF TARA at test time rather than written into the spec, so a reply naming something the database does not hold is caught whatever it says. Four defects, three fixed and locked. EVE-VIS-076 (S1): a streamed agent turn that runs its tools and then FAILS was falling back to the deterministic route, and the abandoned turn's tool list survived the handover to caption the replacement — "I understand you're asking about something related to tara. Could you rephrase that?" under "Checked your recommended practices and the practice you left unfinished", from an engine that ran nothing. EVE-VIS-077 (S2): the streaming turn was handed the deterministic route's TOTAL request budget (12s) as its between-bytes budget; an agent waits out a model round trip after every tool result, so measured over 8 live turns the gaps were 4901, 9029, 5863, 3130, 12168, 4832, 9007, 6763 ms — the matrix fell back on 5 of 24 probes, and after giving the stream its own 30s constant, 0 of 24. EVE-VIS-078 (S1): every agent reply carried "Confidence: 100%" as a full green bar, from a hard-coded confidence: 1 — and it sat under a turn that called no tool and offered the member four favourite meditations, three of which do not exist. AssistantResponse.- confidence is now optional, which made the compiler name the two other places that assumed a number. The fourth is open and honestly so. EVE-VIS-080 (S1): asked what they had favourited, the agent skipped the lookup in 4 of 6 cells and was wrong every time — including a meditation that exists nowhere — while opening with "Let me pull up your favorites". A prompt sentence was added AND MEASURED rather than assumed: 4/20 → 1/18, Fisher exact p = 0.34, wrong answers unchanged at one per arm. That is noise, so it is recorded as guidance and explicitly not as a fix. The honest-failure half is the best news: with Tara really stopped, 6/6 probes had every tool ok:false, no tool note on any of them, no catalogue content invented anywhere, and every reply said so in its own words. My instrument was wrong four times and each is written down: response.text() on an SSE body the page already consumed comes back EMPTY and reported "no tool ran" for four turns whose notes were on screen; draining before the stream closed attributed turn 1's tools to turn 2; location.pathname without search made a real navigation look like a lie, because the agent pushes /domains/tara?path=…; and the first fabrication check asked only "does the answer name what Tara returned?", which the four-item invention PASSED because the one true title was among them. Naming the right thing is not naming only right things. Also found here, owned elsewhere: EVE-VIS-075 (S2, a link followed from part-way down one page lands part-way down the next — both remedies Next's own warning names were tried as counterfactuals and neither changed it) and EVE-VIS-079 (S2, the member reads the agent's internal monologue: "Let me grab that for", then a question the same reply answers for them).)
  • 4.2 Domain read tools — veritas (trending briefs/articles, continue reading, top claims, trending topics, saved articles). _(2026-08-07 s11: none of these tools exists for a V1.0 member, and that is the task's real content. The release cut opens four rooms and defers Veritas and Metis to V1.2; a live member session's authorizedDomains is ["tara","nyx","arete","nisaba"], and agent-tools-release-scope.spec.ts already pins that the model is never SHOWN a deferred room. So the honest 4.2 is not "drive six Veritas tools" — it is what a member who asks for news, fact-checking, saved articles or trending topics actually SEES. Four asks per cell, both themes, three viewports, live model (e2e-inspect/phase-4-2-veritas-deferred.spec.ts): 24/24 probes clean on every lens — no veritas__ormetis\__tool ran, no navigation toward a deferred room, nothing offered, page never scrolled sideways, contrast clean, console clean. One defect, fixed and locked. EVE-VIS-082 (S2): the member was told their PLAN was the reason. Live in 4 of 6 cells — "Your plan gives you access to Tara, Nyx, Arete, and Nisaba", "The only domains available on your plan are…", "it's not a room available in your plan" — and the model was not inventing it: system-prompt.ts instructed "let them know it's not available in their current plan". Veritas is deferred for EVERYONE, so no plan opens it and the member is being pointed at an upgrade that does not exist. Now: "say plainly that it is not part of this app — do NOT say it is missing from their plan". After the change, 0 of 24 replies used plan framing, and the lock asserts both halves (the old phrasings absent AND the new instruction present, so deleting the sentence cannot pass). Mutation-calibrated. My lens was wrong once and it is worth recording why: the offer-detector flagged three HONEST replies, because "I can't take you to Veritas" contains "take you to Veritas" — a lens blind to negation accuses the product of exactly the behaviour it is getting right, and it now checks the 40 characters before the verb. Observed once, did NOT reproduce, and therefore not ledgered: one cream/zoom200 run reported the newest reply entirely outside the viewport (hits: 0, sampled: 0). A clean re-run of both zoom200 cells was 19/19 on all eight probes with healthy geometry (bubble 146px into a 385px transcript, nothing above or below the fold). Per the EVE-VIS-014 lesson a finding that will not reproduce is not a finding — but the transcript geometry is now captured on every probe, so a recurrence arrives diagnosed instead of mysterious. Also found here, owned by 4.4: EVE-VIS-081 (S1) — a member who has never opened Nisaba is told, in the second person, about two notebooks "last updated today", a passage they "were working on", and an offer to "pick up where you left off". The BFF's whole Nisaba adapter is a fixture with the member's id spliced into it.)_
  • 4.3 Domain read tools — nyx (nightly highlights, event detail, continue observation, search objects, logs, saved, reminders). (2026-08-07 s12: seven capabilities named, three reachable — and the assistant's two engines disagree about which. Verified rather than assumed: observation logs has an agent tool and NO deterministic intent; event reminders has a deterministic intent and NO agent tool; event detail and saved objects have an action-router arm no intent can select; continue observation is in neither. So the kill switch (11.2) silently changes what a member can ASK, not just which engine answers — EVE-VIS-087 (S3), table in the spec header. Driven live through the panel (e2e-inspect/phase-4-3-nyx-read-tools.spec.ts), 9 probes × both themes × three viewports, ground truth read out of the BFF's own Nyx adapter routes at the same minute — the sky moves, and a spec carrying its own copy of tonight's moon is a lie by tomorrow. The observation log is asked TWICE with a real write in between (the product's own POST /v1/nyx/observations), so a tool answering from a fixture would say the same thing both times. Three matrices: 54 probes before, 51 after the prompt fixes (3 lost to a hung cell), 54 confirming. EVE-VIS-086 (S2), fixed and MEASURED: asked "What's in the sky tonight?", the agent asked the member where they live and called nothing. Both location parameters were described as "the member's location" while required said otherwise, and a model reads the prose — yet moon phase, illumination and event dates are the same from everywhere on Earth. Six samples cannot tell a change from noise, so it was measured through the BFF's own /turns, 20 per arm: 4/20 looked it up, 8/20 without that session's other prompt change (Fisher p = 0.30 — one arm, not two), 20/20 after the fix; pooled 12/40 → 20/20, p = 1.07e-07, and 0 of 54 in the confirming matrix. EVE-VIS-084 (S2), fixed: the assistant did not know what day it was — the whole clock in the prompt was the word "evening" — so a two-day-old journal entry came back as "earlier today you logged the Andromeda Galaxy". Every tool that returns a record returns an ISO timestamp, so this was never a Nyx bug. 2 of 54 → 0 of 54, six mutation-calibrated tests. EVE-VIS-083 (S2), NOT closed and honestly so: the agent tells members where things live inside a room — "in Nyx under your saved objects list" (there is no such web surface at all) and "from the event card" (true). Being right by luck is the finding. A guideline-11 paragraph was added and measured: 5/54 → 2/54, Fisher p = 0.437. That is noise, so it is recorded as guidance and NOT as a fix. EVE-VIS-085 (S1), open: at 200% zoom four consecutive replies landed entirely below the fold with no jump pill, while desktop and narrow were 19/19 painted in the same run — a 705px transcript absorbs the same shortfall. My first mechanism was DISPROVEN (a doubled stream with the completion-growth ordering passes against the unfixed build), it did not recur in either later run, and the harness now records the panel's own scroll target so the next occurrence arrives diagnosed. EVE-VIS-073 held live: zero wrong moon phases and zero events misdated as tonight across all 159 probes. My instrument was wrong nine times and every one is written into the spec — the object-name extractor read "No, not tonight", "Mark your calendar for August 12th", "Also worth noting" and "Magnitude 3.4" as celestial objects before a Title-Case rule replaced the growing stopword list; a ±80-character window walked across a paragraph break and called a correctly-dated eclipse paragraph a lie; "1 hour 30 minutes" derived from two real timestamps was called an invention; "you can set a reminder from there" was read as the agent claiming to have set one; and a logs-history turn that had fallen back to the DETERMINISTIC engine (the tell was a confidence meter and chips no agent turn produces) was reported as the agent skipping a lookup — the trap now records which engine answered. Also found here, owned elsewhere: EVE-VIS-080 got its sharpest evidence yet (the member's data provably CHANGED mid-conversation and one cell answered "Your log is still empty — no observations recorded yet" without looking, after opening "Let me check again…"), and a navigate to Nyx's events page opened a full-screen overlay ON TOP of the panel so the sentence explaining what just happened was covered — task 4.6 owns that.)
  • 4.4 Domain read tools — arete (active goals, continue plan) and nisaba (daily passage, continue reading, workspace entries, search library) and metis (recommended courses, continue learning, search catalog). (2026-08-07/08 s13: nine capabilities named; the room that answered most confidently was the one making everything up. Driven live through the panel (e2e-inspect/phase-4-4-arete-nisaba-metis.spec.ts), three full matrices — 54 + 47 + 66 probes — across both themes and three viewports, ground truth read out of the BFF's own adapter routes at the same minute, with a real goal and a real notebook written through the product's own routes mid-conversation so a tool answering from anywhere but live rows says the same thing twice. EVE-VIS-081 (S1) CLOSED. Nisaba's whole assistant adapter was object literals in routes/assistant.ts — "your two active notebooks" named after the member's own id, a passage they "were working on" at 62%, and a daily reading that was not the one the room served. The real adapter was already in the process and already honest ([] and null for a member who has done nothing), so the fix was deletion: both Nisaba and Metis now delegate like the other four rooms. Verified live on an account seconds old — "You don't have any Nisaba notebooks or workspaces yet — the shelf is empty" — and the daily passage the assistant names is now byte-identical to the store's. EVE-VIS-088 (S1) CLOSED, and it is what made the Arete half testable at all. Every create the Arete Oshun facade served answered 500 for every member, always: it minted nanoid() ids for uuid columns. Goal, habit, habit completion, streak, journal entry, wheel-of-life check-in. The identical bug had been fixed in eight NATIVE routes, each carrying a generateId() helper whose comment says so in as many words — the facade was never brought along, and the same file already used crypto.randomUUID() correctly two hundred lines up. 4/4 creates 500 before, 4/4 201 after. No test could have caught it: every suite in the package mocks drizzle-orm and runs memory repositories, so the one component that rejects a nanoid was never in the room — which is why the lock checks the UUID shape instead of needing a database. EVE-VIS-094 (S2), fixed and MEASURED, not closed. EVE-VIS-082's own fix traded one false statement for another: its replacement instruction was "say plainly that it is not part of this app", and the model said exactly that — "there's no Metis in this house", "I'm not aware of anything called Metis in the app". Metis is not imaginary; it is a room the product ships, held back from V1.0, and release-scope.ts already carries the member-facing sentence saying so. n=40 per arm through the BFF's own /turns: 29/40 → 11/40 denials (p = 1.0e-04) and 0/40 → 16/40 attributing it to the version (p = 4.66e-06). Real and large — but 11/40 still deny, so the row stays open. EVE-VIS-082 held throughout: 0 of 80 turns blamed the plan. EVE-VIS-095 (S1), open — the sharpest thing in the run. One cell answered "Search the library for something about the mind" with TOOLS RAN: [] and a formatted four-item results list opening "Here's what came up for 'the mind' in the Nisaba library". Three of the four texts do not exist, one of them a book with an author's name on it ("Walking the Suttas: A Lay Meditation Guide (John Peacock)"), and the real translation was re-attributed to a "(Bradley/Müller)" who does not exist. Absence verified against the corpus itself. That is EVE-VIS-080's mechanism with a larger blast radius: a member can go and cite an invented sutta reference. EVE-VIS-080 gains its sharpest evidence yet. With a notebook created between two identical asks, notebooks-real skipped its lookup in 9 of 17 cells across three runs, and three of those told the member their data had not changed while claiming to have re-checked — "empty on both checks", "the list came back empty again", "nothing's been added since we last looked" — while one invented a notebook called "Dhammapada Reflections". And the skips are not random: 0 of 24 bound probes in the first four turns, 7 of 18 at turns 5, 6 and 8 (Fisher p = 1.18e-03). That is the row's own hypothesis about conversation depth, measured. Also filed with evidence and deliberately NOT half-fixed: EVE-VIS-089 (habit completions 500 on a missing unique index — wants a migration), EVE-VIS-090 (nisaba continue-reading has no writer anywhere in the product, which the removed fixture was hiding), EVE-VIS-091 (the agent tools route around every domain's declared assistant read role, in all six rooms; role=assistant has no production consumer at all), EVE-VIS-092 (a deferred room answered from another room with nothing said — 1 of 29, briefly withdrawn and reinstated when the corrected lens found the one real case behind two false ones), EVE-VIS-093 (/v1/nisaba/room serves a module-constant passage unrelated to the store's daily, so the ROOM PAGE and the assistant still name different readings — my fix collapsed three surfaces onto the store and could not reach the fourth). My instrument was wrong five times and every one is in the spec. The plansBlamedFor lens reported Arete's own growth PLAN as subscription-blaming ("Your current plan is 'Walk twenty minutes…'"); claimsInProgress read an OFFER ("Want me to take you back so you can pick up where you left off?") as an assertion, and counted a true "0% done" as invented progress; deferredRoomUnnamed tested for the WORD "Metis" and cleared any reply containing it, which hid the one genuine substitution behind two false ones; and worst, the invented lens inherited 4.3's "no catalogue name contains a comma" — true of a sky catalogue, false of scholarly citation — plus a character class that excluded colons, which is why it reported [] for the fabricated bibliography above. It is rewritten and calibrated against that exact reply as its positive control with four honest replies as negatives. Model deepseek/deepseek-v4-flash-0731, pinned mid-session on the user's instruction. The reason I first gave for pinning it was wrong and is corrected in the harness recipe above: I compared /v1/models summary fields and reported the snapshot as 1.56x cheaper. Priced against the live API instead — per-provider /endpoints, plus usage: { include: true } for the billed rate — a slug on OpenRouter is an auction, and the snapshot is the DEARER of the two ($0.102/$0.180 against the alias's $0.068/$0.137). It stays bound because reproducibility is worth 1.5x on a model already ~100x cheaper than the frontier, which is a decision, not an accident. The real lever turned out to be provider routing, which was not being sent at all: three consecutive calls drew three different providers at three prices and three quantizations. OPENROUTER_PROVIDER_SORT=price now pins both. Phases 0–4.3 ran on the alias under default routing, so their model label is weaker than it looks — that is recorded rather than retrofitted.)
  • 4.5 Do-tier mutating tools with confirm cards on the MEMBER panel: tara add/remove favorite, veritas save/unsave + follow/unfollow, nyx log observation + set reminder — for each: card copy names the exact action, Approve executes + confirms visibly in the domain UI, Decline leaves no trace, expiry (>5min) renders honestly. (2026-08-08 s14: DELIBERATELY LEFT UNCHECKED. Seven defects found, all seven fixed and mutation-locked, every fix verified live end to end — but the six-cell visual matrix has ONE cell complete, and this file's rule is both themes x three viewports. What was established first is the task's own content: eight capabilities are named and TWO are reachable by a V1.0 member. tara_add_favorite and nyx_log_observation are bound; Veritas' save/follow are bound to a room deferred to V1.2; and remove-favorite, unsave, unfollow and set-reminder are implemented and wired onto the assistant's own adapter surface in buildAssistantDomainAdapters and bound to no tool (EVE-VIS-101, S3 — S3 rather than S1 because the agent is honest about all four when asked live, verified: "I can add favourites for you, but I don't have a way to remove one from this side", and the row was still there). EVE-VIS-096 (S1): nyx_log_observation had never worked for anyone, on either engine. The Nyx store requires conditions — deliberately, so a sky nobody described is never recorded as 'clear' — and the tool's schema had no such field. Live: the member said "Log that I observed Jupiter tonight — the sky was clear", the card parked, they confirmed, and the BFF answered 502 … got 'none'. The failure arrives after the member has authorised it, and every existing test asserted the handshake and stopped one line short of running the parked closure. The deterministic engine had the same hole from the other end: its planner collects the member's reported sky into the action params and its executor dropped it. EVE-VIS-097 (S1): the Moon was not in the celestial catalog, though every "what's in the sky tonight?" answer LEADS with it — so asked to log the object it had just described, the agent said "I couldn't find the Moon in the sky catalog this time". The new lock pins the RULE (every name in a highlight's objectNames must be searchable) and caught the Sun on its first run — named by every solar eclipse, equally missing. EVE-VIS-098 (S2): the card read Add meditation "550e8400-e29b-41d4-a716-446655440001" to your Tara favorites. under a model sentence saying "Calm Your Mind" — the unreadable one being the control that performs the write. Names are now resolved out of the domain, so an invented id refuses BEFORE a card is shown. EVE-VIS-099 (S2): the provenance line called a parked write a lookup ("Checked your … saved practices" for a favourite not yet saved, and still there after Decline) — the module's own rule already existed and the four write tools had been left out of it. EVE-VIS-100 (S2): the card carried no deadline, so it offered Confirm forever and reported expiry in our words; measured with a REAL 5m20s wait (registered 09:03:26Z, tapped 09:08:46Z, 404 assistant_action_not_pending, nothing written). EVE-VIS-102 (S2): the Confirm button measured 2.09:1 in dusk — accent text on a 9% tint of the same accent — in 6 of 6 cells; fixed by composing with EVE-VIS-025's existing --assistant-on-accent machine rather than writing a second one. Verified live after the fixes, through the BFF with the panel's own client capabilities: Jupiter and the Moon both log (CONFIRM → 200, rows in the journal with conditions: clear), the Tara card reads Add "Calm Your Mind" to your Tara favourites., and a member who has not described the sky is ASKED ("how was the sky when you spotted Saturn? Clear, partly cloudy, overcast, or washed out by city light?") with no card parked and no row written. The one matrix cell that completed (cream/zoom200) is clean on contrast, covered-controls and overflow, with the card naming the row, the row landing, Decline leaving 0 observations, and the expiry note reading "That confirmation is no longer open — these expire after a few minutes. Nothing was changed." What remains, precisely: four other cells, and EVE-VIS-103 (S2), which the second measured cell turned from a zoom-only curiosity into a general one: at 1470x900 the card's Confirm renders at y=898.5 with a height of 31.6px — 1.5px visible — and a second card's buttons at y=952.4, fully below the viewport, so the member is told to tap a control that is not on screen. cream/desktop and cream/zoom200 are otherwise CLEAN on every lens (contrast 0, covered 0, overflow 0) with all behaviour correct. My first mechanism for it was disproven and the fix reverted rather than shipped — the card sits below the transcript's end sentinel that scrollTranscriptToLatest aims at, so a scroll-the-card-into-view effect looked certain and moved the number not at all (898.6 → 898.5). Next step is one geometry capture (panel, body scrollTop/scrollHeight/ clientHeight, card, composer), not another theory. Two harness lessons are written into the spec header: Send is disabled by an EMPTY composer as much as by a turn in flight, so "wait before typing" can never come true; and a stash-and-compare across a dev server started before both arms is worthless — it produced a fully-worked false conclusion that a regression was mine.) (2026-08-08 s15: CLOSED. All six cells green — offscreen 0, covered 0, contrast 0, overflow 0 — and the console lens passing for the first time in this task's life. Every Confirm button lands inside the viewport in all three card positions of all six cells (desktop 750-783, narrow 715-748, zoom200 321-354). Behaviour holds throughout: the card names the real meditation, Confirm lands the row, Decline leaves zero observations, and the expiry note reads "That confirmation is no longer open — these expire after a few minutes. Nothing was changed." EVE-VIS-103 CLOSED, and the residual was not what the row said it was. Three instruments in sequence: an eight-stage deterministic replay (two cards from one turn, the spec's own .first() tap) came back clean at every stage and killed the "a second card is left parked" theory; instrumenting the live cell pinned the third card at 909-988 against the panel's 879 bottom; and a per-animation-frame timeline over three live cards showed the ENDPOINT was never wrong — every card ends at 801, 12px above the composer. What was wrong is that each one ARRIVED at 996-1050 and took 1053ms and 1472ms to come into view, live and tappable, while the reply saying "confirm it below" was still being written. pendingActions was missing from the follow effect's dependency array: a card lands on turn.ui, independently of messages, so nothing moved the transcript when one appeared. After: 103ms, 176ms, 46ms — the scroll animation itself. Three harness defects, each of which made this task measure fiction. ask() counted .assistantMsg--assistant, which the typing indicator also matches, so it returned the instant the send went out — that is WHY the third card was first filed as a permanent position, having been measured one second into a transient window. The console lens asked for consoleTrap.errors, which the trap does not publish, so it had never once reported a message; corrected, it immediately found EVE-VIS-109 (Next overriding the app's own scroll behaviour on every navigation for want of data-scroll-behavior). And EVE-VIS-104: the evidence directory lived inside outputDir, which Playwright DELETES at the start of every run — watched happen, two cells written at 12:37 and 12:39 and gone after the next playwright test. That is the likeliest reason two sessions reported "ONE cell complete": the others had been measured and then deleted, which reads identically to never having run. EVE-VIS-106 (S1) found here, fixed, and it belongs to 4.6. Three cells wedged with the harness waiting three minutes on a composer that no longer existed; the snapshot read breadcrumb Home / Tara with an "Open AI assistant" launcher and no panel. At phone width the agent's own navigation closed the assistant and took the conversation with it, because the dock persists its open state and the overlay did not. Reproduced by scripting response.navigateTo so the model cannot vary: desktop true / 2 / 1, narrow false / 0 / 0; after, narrow reads what desktop reads. EVE-VIS-107 (S1): the domain-UI half of this task's own sentence was false. "Approve executes + confirms visibly in the domain UI" — the matrix verifies the write through the BFF, so I opened the screen. TaraFavorites had no data source at all: five invented practices, invented ratings and play counts, a savedAt recomputed from Date.now(), and ~70 tests asserting all of it. A member confirmed the card, the row landed, and Favorites showed them five things they never saved. Rewritten onto the real rows, the three unsourced fields deleted rather than placeheld, verified live in four cells with an empty-shelf control. EVE-VIS-108 (S1) fell out of that fix failing two of four cells: resolveBffAuthToken() falls back to a fixed sub: "u123" token whenever the session accessor is late, so a first-paint fetch is answered with another member's rows — 200 [], indistinguishable from an empty shelf. Only this call site is fixed; TaraSurface has the same shape and is ledgered. Still open and deliberately not fixed here: EVE-VIS-105 (S3) — the card sits below the "Continuity journey map" disclosure and the relevance bar, so unrelated chrome separates the sentence from the control it names; the fix moves a JSX block the EVE-VIS-103 scroller now anchors on, so it wants a full re-measure at the phase boundary. EVE-VIS-101 (S3) unchanged: four capabilities implemented, wired onto the adapter surface, bound to no tool — and the agent is honest about all four. asksNeededForCard was 1 in five cells and 2 in dusk/narrow, which is EVE-VIS-080's family and is recorded rather than retried away.)
  • 4.6 navigate: every domain destination via conversation — router push (no reload), panel stays open with context, breadcrumb/url correct; invalid destination refused gracefully in-chat. (2026-08-08 s14: SOURCE GROUNDWORK ONLY — no browser run, and the checkbox needs one. The box is inherently visual (router push vs reload, panel state across a push, breadcrumb) and this box was actively paging (8,661 pageins in 20s, 84 MB free pages) after the 4.5 matrix, so starting a Chromium sweep would have manufactured hung cells rather than findings — the 4.5 run had already shown exactly that, six cells green then one green and five hung with no app change between them. What was established without a browser, for the next session to start from: (a) the destination space is narrower than the task text implies — navigate's domain enum is built as NAVIGATE_DOMAINS.filter(d => authorized.has(d)), so a V1.0 member's model is offered four rooms, not six, and the run refuses an unauthorized domain a second time. (b) navigate does NOT validate that the destination exists. It checks the domain is authorized and that path starts with /, then acknowledges — so the model can promise any path it likes and the tool agrees. Task 4.6's "invalid destination refused gracefully" is therefore about a refusal that does not currently exist for paths, only for domains; whether a member ends up somewhere wrong depends on what each room does with an unknown path, which is the browser half. (c) the URL is not the destination. buildHydratedWebDomainRoutePath turns {domain, path} into /domains/<domain>?path=<encoded>, so every navigate lands on the room route with the destination as a query param — which is why 4.1's instrument was wrong when it read location.pathname without search and called a real navigation a lie. The param is NOT inert: the page passes it to DomainRouteExperience as hydratedPath, which threads it into the room's own rendering and into the route title. Whether each room then actually opens the named thing is the first question a browser pass should answer. Already-known inputs to this task, from earlier sessions: the agent fires navigate unprompted on a plain recommendation ask and says "I've opened Calm Your Mind in Tara for you" (4.1/4.5 recon), and a navigate to Nyx's events page opened a full-screen overlay ON TOP of the panel, covering the sentence explaining what had just happened (4.3). Both belong here.) (2026-08-08 s16: CLOSED. Six cells green — cream and dusk at 1470, 390 and 735 — with five destinations driven through the panel in one conversation per cell, and both of the earlier sessions' inputs turned out to be the same defect. Per hop: the push is a push (a witness planted before it survives in every cell, so no reload), the URL is exactly buildHydratedWebDomainRoutePath's output, the breadcrumb names the room, the room paints its real screen (viewWitnessPainted true for all four authorized destinations), and the transcript grows 1→5 member bubbles across four consecutive navigations without resetting. Contrast, overflow and console are clean. Four defects found, three fixed. EVE-VIS-112 (S1) is 4.3's "full-screen overlay on top of the panel", and it was not one room — three of the four authorized rooms buried the assistant: elementFromPoint at the composer's centre returned div[Course browser], div[Habits] and canvas, with the reply that said "I've taken you there" underneath. 25 inline position: fixed; inset: 0 screens in Tara and Arete plus 9 CSS containers in Nyx, none aware the dock exists; the shell now publishes --shell-assistant-inset (beside EVE-VIS-026's --shell-bottom-chrome) and they clear it. Only the DOCK presentation was affected — the four cells below ASSISTANT_DOCK_MIN_WIDTH were clean throughout, because there the assistant is the overlay and is itself on top, so a phone-first look would have called this fine. EVE-VIS-113 (S2): opening the assistant deleted the breadcrumb at 1470 and 1600 (trail 248.5→0) while its labels stayed correct in the DOM — the actions row is flex-shrink: 0 and the trail absorbed the dock's whole 380px. The bar had a compact presentation for exactly this and was asking the VIEWPORT. EVE-VIS-114 (S3): geolocation=() blocked the app's own origin, so Nyx's sky map logged a violation on every visit and drew New York "(default)" — the same bug as EVE-VIS-041's microphone, and this repo's own lock asserted it. EVE-VIS-115 (S1) is 4.1/4.5's "I've opened Calm Your Mind in Tara for you", and s14's open question answered: the tool accepted any path starting with /, and no room fails on an unknown one — every classifier falls through to its HOME view, silently. The tool's own example, /sessions/med-001, matches nothing Tara implements. Fixed with one registry in @oshun/navigation that is load-bearing on both sides — navigate describes and REFUSES from it, and each room's screens hang off its keys, with a 28-test parity spec failing in either direction. Live after: that same ask now goes to /meditate/session/<real id> and the sentence is true; "take me to Nyx's telescope rental store" gets no tool call and an honest refusal with four real alternatives. Two adjacent findings ledgered, not fixed: EVE-VIS-116 (S2) — four Nisaba CARD destinations name screens Nisaba does not open, found by checking the formatter's path literals against the new registry, and closing it is a card-surface decision for the Phase 4 boundary; EVE-VIS-117 (S3) — the agent answers "can you?" from the docs corpus rather than its tools, so it denied Nisaba's lexicon view, which the room implements. Also honest about the environment: the sweep needed a pro member (OSHUN_DEV_SIGNUP_PLAN=pro) or Arete answers with its paywall, and 12 pre-existing failures in src/components/domains/ (assistant-handoff and concept-graph tests) were confirmed red on a stashed tree before and after.)
  • 4.7 highlight → AnchorSpotlight: drive a highlight on every seed anchor (all registered anchors — enumerate from assistant-anchors.ts); spotlight geometry hugs the target at all three viewports, scrim contrast right in both themes, dismiss paths work, no scroll-position bugs when target is off-screen. (2026-08-08 s16: CLOSED. All seven registered anchors driven through the panel's own turn.ui highlight intent, six cells green — cream and dusk at 1470, 390 and 735. Contrast was the one thing already right: the ring measures 4.01:1 on the scrim in cream and 14.32:1 in dusk (both over the 3:1 non-text minimum) and the note 16.13:1 / 10.7:1, with a clean console in every cell. Three defects found and fixed. EVE-VIS-118 (S2): shell.primary-nav had not rendered for anyone in months — it was stamped on ShellRouteNavigator, the pill nav the sidebar replaced, which nothing imports; the coverage guard greps SOURCE, so it stayed green while the anchor was absent from every page. Two more anchors are absent by design (the palette exists only while open, the floating launcher only on shell-less pages) and both claimed routePrefix: '*'; the registry now carries an explicit availability field. EVE-VIS-119 (S2): two of the four reachable anchors are <section>s taller than the screen (home.daypart-rail 764×1780 at 1470×900, 362×2636 at phone width), so the cutout ran off the top and bottom at once and the scrim dimmed nothing — the member was told "that's here" and saw no change. Now clamped, with the clamp deliberately narrow: two earlier cuts each broke a working case (the composer trimmed 57→37px, the phone nav clamped narrower than itself). EVE-VIS-120 (S2): Escape dismissed the spotlight AND closed the conversation, in all six cells — three window listeners answering one press. The first fix, a shared attribute the other handlers checked, was NOT enough and the matrix said so (three anchors kept the panel, two lost it); the mechanism that works is a capture-phase listener that stops propagation. Two harness lessons worth carrying: a lock pointed at a case that never failed passes against the broken build — this one's first two versions did exactly that — and a SEEDED ACCOUNT accumulates state that changes what Home renders, which produced an hour of "the fix does not work" against a fix that was present and correct. Recorded, not fixed: the spotlight blocks input to everything including the element it points at (22 of 24 hit-tests returned the scrim), which is the dismiss-on-click design — but the component's own comment claimed it "never blocks input", and that comment is now honest.)
  • 4.8 read_page client tool: paused-turn round trip visible states (what does the member see while the client tool runs?), timeout rendering, and a page with data-assistant-private content — verify exclusion. (2026-08-08 s17: CLOSED. Six cells green — cream and dusk at 1470, 390 and 735 — against the real BFF, the real model and the real bridge; nothing scripted. The paused window is invisible from outside the page (page.on('response') fires when the SSE response STARTS, not when a frame six seconds in lands), so e2e-inspect/support/turn-wire-observer.ts TEES the real body and timestamps every frame while a rAF sampler records what the transcript paints at the same instants. Answered: the round trip is imperceptible — 0.1ms from frame to POST, a 6–7ms paused window (136ms in one cell), 202, ok: true, and "Checked the page you are on" under the reply, in all six cells on the first ask. Timeout: honest, and so is the provenance line. With the POST aborted, the server's own 10s bridge timeout resumed the turn (9943–10007ms measured); the member saw ONE unbroken state for it — typing dots labelled "Lilith is writing a reply", composer live, Send disabled — the reply said "The live page read didn't come through, so I'll go by what the shell shows me…", and the "Checked …" line was ABSENT, so a failed lookup is never described as one (the EVE-VIS-099 shape, holding). Three defects found and fixed, two ledgered. EVE-VIS-121 (S2): data-assistant-private marked nothing — read by three collectors, written by nobody, 0 subtrees on every route measured, including the notebook page whose own header says "private". Now a registry (ASSISTANT_PRIVATE_REGIONS) with a DOM guard, because a grep could not tell the difference — EVE-VIS-118's family exactly. EVE-VIS-122 (S1): the selection channel ignored it, so 167 characters of the member's margin question went out on the wire while the Ask-Lilith chip stood down and implied it would not; and the chip's own check read only anchorNode, so a drag ENDING in private text walked past it. EVE-VIS-123 (S2): read_page returned the ambient block byte-identically while promising "visible controls" — it now returns the controls by role, name and state, and never a value. Two adjacent findings ledgered, not fixed: EVE-VIS-124 (S2) — seven of ten Nisaba depth pages are wider than a phone (/nisaba 1285px against 390), a room-layout class for the Phase-12 sweep, still MEASURED here and written into this task's evidence rather than suppressed; EVE-VIS-125 (S3) — the two journey-inventory gates cannot both be green, and the freshness one was already red on a clean tree, so the regeneration was reverted rather than destroy the parity golden. Three harness lessons worth carrying: (1) .assistantMsgBubble:not(.assistantTyping) matches the STREAMING bubble from its first delta, so counting bubbles says "done" mid-reply — the first run concluded that a SUCCESSFUL read produced no provenance line and a FAILED one produced "Checked the page you are on", and both readings were the same note, one turn late; (2) a privacy assertion NEEDS a positive control, because "the private text did not arrive" is satisfied just as well by a selection that collapsed before Send, and it does collapse intermittently while the composer holds focus; (3) a needle taken from text-transform: uppercase text can never appear in Selection.toString(), and one sitting past the collector's 1,000-char cap is truncated away — each makes not.toContain pass however badly the product leaks, and the mutation pass is what said so.)
  • 4.9 search_docs: grounded answer with source naming; member corpus boundary (a doc only in the FULL corpus must not leak to a member session); "docs don't cover it" honesty; grounding chip UI. (2026-08-09 s17: CLOSED. The boundary question turned out to be the wrong shape: nothing needed to leak FROM the full corpus, because the MEMBER corpus was already the engineering documentation center. EVE-VIS-126 (S1), fixed: 317 pages / 6,485 chunks of docs-center/** were being served to members — every V-series product page (Ariadne · engine — the self-authoring detective universe), the repo map, 338 API and 330 contract references, 60 Rust-crate pages, and ten start-here lenses for Engineer, SRE, Exec/Investor and Security. Live, a member asking how to save an article was told "the library docs that come up are internal engineering material". The rule is now a positive opt-in in the builder, nothing in the estate declares itself member-facing, so the member corpus is EMPTY and an empty corpus fails closed — search_docs is not offered to member sessions at all (user decision, 2026-08-09). Verified live after restarting the BFF — it caches the corpus per scope, so a restart is required; its env was captured to a 0600 temp file and it was relaunched identically rather than guessing the pinned secrets. MEMBER: search_docs no longer appears on the wire at all, the two asks that used to reach it now run only domain tools, the "internal engineering material" phrasing is gone, and "tell me about Ariadne" returns "that's not a universe I have on record". ADMIN: still offered, still grounded — "The page is Repo map — the system in 30 minutes (docs-center/reference/repo-map.html)" — and still honest where the corpus is silent: "The docs don't cover a refund policy… I can't quote those as policy." Six browser cells green (e2e-inspect/phase-4-9-search-docs.spec.ts, cream and dusk at 1470, 390 and 735), reading the WIRE for the absence of search_docs rather than inferring it from a reply — a model that merely chose not to mention Ariadne looks identical to a tool that was never offered. EVE-VIS-127 (S2), ledgered not fixed: agent turns set neither grounding nor confidence, so the grounding badges and the confidence bar are dead for every streamed reply — and since streaming is the default, for every real member reply. Wiring them means putting evidence vocabulary on "two lookups succeeded", which is what EVE-VIS-078 declined to do for confidence, so it belongs with the 11.7 disclosure pass. The grounding-chip leg is therefore a MEASUREMENT rather than an inspection, and deliberately so: the spec asserts groundingRows === 0 and confidenceBars === 0 with a message telling a future session to update the expectation and inspect the thing once it renders. Two findings worth carrying. The test that should have caught EVE-VIS-126 asserted the same premise the defect rested on — every result link must START with docs-center/, under the heading "member-safety boundary" — so a prefix stood in for an audience and the guard passed for years; it is rewritten in place with the disproved premise recorded rather than deleted. And the overlap lens reported the newest reply as covered by the composer in ONE of six cells, which is the signature of a race rather than a defect: inspect's 600ms settle is not enough for the panel's smooth scroll at a 450px-tall viewport, and measured AT REST (probe-4-9-composer-overlap.spec.ts, scripted so the reply LENGTH is controlled instead of whatever the model wrote) the bubble ends 57.3px clear of the composer and hit-tests to its own text in both themes. The spec now waits for the transcript's scrollTop to stop changing before it looks.)
  • 4.10 Audit tools in conversation (deep pass is Phase 7): begin / status / mark visited / mark skipped-with-note — including the model REFUSING a noteless skip gracefully. (2026-08-09 s18: CLOSED, with the graceful refusal turning out to be the whole task. EVE-VIS-128 (S1), fixed: a skip reason the member never gave, recorded as theirs. The store refuses a noteless skip and its comment says why — "a skip without a reason is exactly the silent gap audit mode exists to prevent" — but the note is a TOOL ARGUMENT, so the model writes it, and asked only "Skip the next flow." deepseek-v4-flash-0731 satisfied the guard with note: "Member requested to skip this flow." and told the member "Skipped Switch between domains. Coverage is now 1 visited, 1 skipped." The row is still in the durable snapshot (customer-eve-4-10-dusk-desktop-779300), the only one of 21 marks across 11 cells whose note is not a reason. Presence is what the store can see; PROVENANCE is now checked at the agent boundary (audit-skip-reason.ts) — a note must say something beyond the request once the skip vocabulary is stripped, AND contain a substantive word the member actually typed, from their last three messages only (unbounded, an audit conversation leaves "walked" and "visited" lying around for an invented reason to grab). Live after the fix, every recorded note is the member's own sentence: "On a shared computer and would rather not open personal content in public." EVE-VIS-129 (S2), fixed: builder mechanics in the transcript — raw catalog ids in 3 of 6 cells ("Done — shell.navigation.tabs is marked visited") and internal retry narration in 3 of 6 ("That flow id didn't match the active run — let me pull the exact one"), because the agent fumbles an id it retyped and audit_mark threw a bare audit_flow_not_in_run. resolveFlowRef now resolves by exact id, punctuation-insensitive id, or the flow's exact TITLE (745 of 747 titles are unique; the two that share one refuse rather than guess), and a genuine miss returns the correct id and title IN the error, so there is no second failure to narrate. EVE-VIS-130 (S1), fixed — the server now checks the claim. Replies that claim a skip the run never received: 3 of 11 cells, each confirmed against the snapshot, two with tools: [] behind the claim. This is EVE-VIS-080 instantiated (a claim about member-owned data no tool supports), the conduct block already asks for it in two places, and its own comment says a request is not the mechanism — so the mechanism got built for the one kind of member data whose truth is a handful of integers. audit-claim-check.ts parses the progress numbers a reply asserts and the completed-mark phrasing, compares them against the run, and flags only an OVERSTATEMENT (an understatement cannot mislead a member into thinking work is done, and future-tense arithmetic would otherwise be corrected for nothing); the route appends one member-language correction to the authoritative transcript and streams it, the same property the post-safety override beside it already uses. The fix is what makes the control possible: the occurrence is intermittent, so no assertion about the MODEL could be anything but flaky, and the assertion is instead that the SERVER caught it — deterministic. The four near-misses matter more than the catches: "I can mark it skipped once you tell me why" and "Shall I mark it visited?" are the CORRECT behaviour, one modal away from the defect, and the first draft flagged two of them. Three method notes. (1) The old assertion was skipNoReason.tools does NOT contain audit_mark — an assertion about the WIRE, which passed six cells in a row while the fabricated row sat in the database; it now reads the ROW (GET /v1/assistant/audit-runs/current), so a call that is refused is fine and a row that appears is not. (2) The transcript guard was a list of snake_case identifiers and returned [] in all six cells while machine vocabulary was on screen in five — the model paraphrases identifiers ("flow id") and prints their VALUES; it is replaced by four patterns taken from the measured transcripts, each with a POSITIVE CONTROL asserting the detector fires on the real leaking sentence. (3) The fabrication detector is numeric (a stated skipped count above the run's, or "marked X as skipped" with no audit_mark) because a first version that looked for the word "skipped" flagged the polite refusal itself. Two harness repairs found on the way: the coverage read needs the seeded session's Bearer token, not the cookie — a cookie-only fetch 401s and would have compared every assertion against null; and when the provider read-times-out the panel falls back to the buffered /message endpoint, which the wire observer does not tee, so the old wait sat for its full 180 seconds and reported a predicate timeout instead of a provider failure — the fallback is now detected from the network (the DOM version is impossible: .assistantMsgBubble:not(.assistantTyping) matches the STREAMING bubble from its first delta, 4.9's landmine, and using it ended the wait mid-turn) and a cell that took it fails saying it measured nothing. Also fixed: 4.9 left the BFF route suite red. A third test shared EVE-VIS-126's disproved premise — it asserted the MEMBER toolset advertises search_docs — and is now two tests: the mechanism on the builder session that has a corpus, and the member half asserted on the wire. Six browser cells (cream and dusk at 1470, 390 and 735, five live turns each), five-lens clean, console clean — five green on the verifying run and dusk/narrow red, green on an immediate re-run (s19 checked the claim rather than inheriting it; the failing run's output was not captured, so what failed there is not known and is not claimed). That is the residual shape of this harness: five live model turns per cell against a provider that read-times-out, and the fellBack guard exists precisely so such a cell says it measured nothing instead of failing an assertion about the product. 32 new unit assertions across audit-skip-reason.spec.ts / audit-agent-tools.spec.ts / audit-store.spec.ts plus three route rows, and five mutation controls including one that proves the route row is the only thing locking the wiring.)

Phase 5 — Tours: every curated tour, every player behavior#

  • 5.1 Walk EVERY curated tour end-to-end visually (enumerate the catalog in tour-plans.ts — shell-orientation, home-daily-flow, assistant-orientation, plus any added since): each step's route push, anchor wait, spotlight fit, narration copy, Back/Next/End, arrow keys, Esc. (CLOSED 2026-08-09 s20. s19 walked all four tours and fixed EVE-VIS-131…135, but left the checkbox open and the walk had never looked: the matrix asserted geometry, copy and controls numerically and wrote no frame at all — which is the exact blind spot this whole file exists to correct. Adding the frames found the rest. The harness was passing a blank screen. spotlightFitVerdict compared the ring's rect to the anchor's and nothing else, and for an OVERSIZED anchor the clamped ring is a viewport-centred box that satisfies every clause no matter where the anchor is — so with shell.domain-switcher 11,495px down Home and home.daypart-rail 1,731px down, the step was declared settled at scrollY 0, every copy and control assertion was made mid-flight, and the evidence for 2 of the 11 steps was a uniform sheet of paper colour Chromium had not finished painting. Two clauses close it — the anchor must be ON SCREEN and the ring must OVERLAP it — plus a settle check that waits for the smooth scroll to stop, and three positive controls built from the exact readings that fooled the old detector. EVE-VIS-136 (S2), fixed: Home's domain launch cards picked a three-column template from the WINDOW's width while living in a 564px rail, squeezing the narrative column to 117px (77px at laptop) — two words to a line, and the launch chip wrapped onto three lines inside a pill. A media-query decision about a container-query problem; that template could not fit until the window passed ~1700px, so it was effectively unreachable. EVE-VIS-137 (S1), fixed: the domain switcher tiles — the thing shell-orientation step 2 spotlights — were painted with the same token that coloured their text, so domain NAMES measured 1.04–1.06 in cream and 1.05–1.14 at the pale end in dusk, and taglines 1.64–2.00. EVE-VIS-025's lesson in a second place. Fixed on the tile (a fixed navy, text inheriting its foreground) rather than the text, because these tiles were always meant to be dark — the rule already said color: #f8fafc. Worst ratio after: 7.07, everywhere, both themes. EVE-VIS-138 (S2), fixed: the narration card was fixed at bottom: 24; left: 50% whatever it pointed at, so it covered 76 of the phone navigation's 96 pixels while reading "These tabs are your main destinations" — 6 of 6 cells. It now tries below → above → right → left, and where a spotlight fills the viewport it says side: 'overlapping' instead of going off screen — a claim the harness re-checks against the rects, so the attribute cannot excuse a placement bug. Three method notes. (1) Two of my own measurements were wrong in the passing direction first: the contrast probe understood rgba( but not color(srgb …), so it silently read ONE of each tile's two gradient stops and reported it as the whole gradient — every row came back 1.04 and the dusk tile's readable white heading looked as broken as its illegible tagline; and comparing against the gradient's worst END rather than the text's own POSITION would have aimed the fix at the wrong half. (2) A frame is not proof either: from the full-page screenshot I read the composer's spotlight as offset from its input and the cream tiles as light-on-dark, and both were wrong — probe-5-1-composer-spotlight and a 3× zoom settled them. Look, then measure what you think you saw. (3) The [data-home-veritas-briefing] leg of viewport-fit.spec.ts fails before it reaches Home's domain grid, verified identical with this session's changes stashed — stale and Veritas-scoped (V1.2), so 136's lock is its own test rather than a line inside a red one. All 9 cells green (6 theme×viewport walks + 3 detector controls), 11 steps each, frames written to .evidence/phase-5-1-curated-tours/shots/ and reviewed.)
  • 5.2 Honest-skip behavior (post-1.5 dialog fix): force a missing anchor (navigate mid-tour); the skip explanation is legible, truthful, and the outcome→replan loop produces a sensible adjusted tour (observed live 2026-08-04 — now verify it VISUALLY reads well). (CLOSED 2026-08-09 s20. Forcing the miss the way the task asks — a step declaring route: '/explore' with an anchor that only exists on Home — found that the skip could not be reached at all. EVE-VIS-139 (S1), fixed: a tour did not survive its own route push. ShellLayout is rendered inside each page rather than in a shared App Router layout, so a client-side navigation unmounts AssistantPanel and mounts a fresh one, and the tour lived entirely in that component's useState. Measured: at t=967ms after Next the URL was /explore, the overlay was gone, and it never came back — so every multi-page tour died at its first navigation and the route field on tour steps was inert. 5.1's spec had recorded the neighbouring symptom and deferred it to 5.3; it is answered here. The transcript has always survived that unmount through the panel's sessionStorage snapshot, so the tour now travels in the same envelope, written synchronously on each step change because an effect scheduled for the next commit is a race against a navigation that has already started. Then the fix produced two of its own, both caught by reading the sentence the member sees. EVE-VIS-141 (S2): the tour survived and its HISTORY did not — "0 of 2 steps shown" to someone who had just watched one (and a second bug underneath it: the panel's persistence effect wrote the tour without the progress, silently overwriting the synchronous write). EVE-VIS-142 (S2): with the tour alive long enough to contradict itself, step 2 located its anchor on the page it was LEAVING — router.push is async — marked itself shown, then missed it on arrival, telling the member "2 of 2 steps shown. 1 step could not be shown." The player now measures nothing until the route it asked for is the route it is on, and the deadline starts on arrival. EVE-VIS-140 (S2), fixed: the re-plan message — "[tour-player] … these steps could not be shown: home.daypart-rail (route /explore)…" — was rendered as the MEMBER's own words, undoing EVE-VIS-133 one statement later. Its comment defended its visibility, which was right, without noticing it had also chosen its author; sendMessage now takes showAsMemberMessage: false and the member gets a system line in their own language. EVE-VIS-144 (S2), fixed: when the agent turn for that machine-authored message failed, the deterministic fallback answered it with "I understand you're asking about something. Could you rephrase that?" — asking the member to rewrite a sentence they had never seen. Confirmed at the route level in the BFF log (/turns then /message on the same session) and from the transcript alone (agent turns set no confidence — EVE-VIS-127 — and that reply had a bar). A turn the member did not write no longer reaches that route. EVE-VIS-143 (S3), fixed: the skip copy named a DOM node ("the element is not on the current page"); it is now "I couldn't find this one on the page you're on, so I've skipped it." Three of my own measurements were wrong in the passing direction first, which is the running theme of this file: the scripted double answered the re-plan with the SAME tour intent, so ending the tour restarted it and looked like a product defect; waiting for the reply BUBBLE returned instantly because a streaming bubble exists from its first frame, so two cells recorded an empty reply as a pass; and waiting only for reply text failed a cell for behaving correctly, because a successful re-plan ACTS — it starts an adjusted tour and unmounts the panel — and a failing one now says so honestly. The wait accepts all three endings and records which. The live loop reads well, on deepseek-v4-flash-0731: "No problem — that step pointed at the Home journey rail, which only lives on the Home page, so let's rebase the tour from where you are in Explore" and "The short detour onto Explore broke the tour — the next stop was on Home, which isn't visible here. Let me replan from where you actually are." Both diagnose the cause correctly in member language; both announce the re-plan rather than starting one in these runs, while an earlier run did start one (recorded, not asserted — that is model behaviour, not a product contract). 8 cells green (6 scripted skip cells across both themes × three viewports, 2 live re-plan cells), five-lens clean, console clean, narration contrast 16.13.)
  • 5.3 Tour + panel interplay: panel hides during tour, returns after, transcript intact, no focus loss; starting a tour from a conversation mid-scroll. (CLOSED 2026-08-09 s20. Three of the four hold: the panel steps aside for a shell-hosted step and comes back, and the conversation returns intact — compared by message TEXT in order, not by count, because a count survives a transcript that was rebuilt with different content. EVE-VIS-145 (S2), fixed: focus was dropped on the floor, twice. Removing the element that has focus sends focus to document.body, so with the composer deliberately focused first, the first step measured body "Skip to main content…" in all six cells — reaching "Next" meant tabbing through the entire shell — and after the tour the panel came back and focus did not. The player now focuses its card on mount (the card, not its button: it carries aria-label="Tour: …" and the live region that reads each step) and the panel's "focus the composer when you open" effect now also runs when a tour ends, since from the keyboard's point of view that is opening. The mid-scroll leg is where I was wrong twice, and it is the more useful half of this task. First lens: I measured .assistantMessages, which never scrolls — .assistantPanelBody is the single scroller (EVE-VIS-022 says so in the CSS) — so scrollHeight === clientHeight guarded the whole scroll assertion out of existence. Second: with the real scroller, I asserted "the newest message must be visible" and got a 958px failure, wrote a settle-loop fix for the follow effect, and only then checked the premise against a plain close-and-reopen — where a healthy panel measured 548–1016px short of the scroller's bottom BY DESIGN, because the continuity rail and diagnostics sit below the transcript. The 958px was the panel obeying its own hold contract: the cell scrolls up on purpose, and EVE-VIS-001's rule is that a member who scrolled up keeps their place and is offered a way back. The speculative fix was reverted. What is asserted now is the disjunction that actually protects the member — after a tour the transcript is either at the newest message or held WITH its jump affordance, never stranded above it with no way back — plus a separate unheld pair that asserts the strict version where it is the right question. That unheld cell was intermittent at a fixed 1.5s wait (twice green, once red); it polls the invariant now, bounded, because an intermittent control teaches the next session to re-run rather than to look. Observation recorded, not fixed: at 1470px the hold survived the tour in every run; at 390px one run released it and followed to the latest. Both leave the member somewhere they can act and I have no evidence either is wrong. 8 cells green.)
  • 5.4 aria-live narration + reduced-motion: tour with a screen reader active (VoiceOver spot check), spotlight transitions under reduced motion. (CLOSED 2026-08-09 s20. Narration is clean: one polite live region per step carrying the catalog's own sentence, one update per step change, and the card has an accessible name ("Tour: Let Lilith show you the house"). VoiceOver itself is not scriptable from this harness, so what is measured is what VoiceOver reads — which regions exist, what they hold, and whether a step change produces one update or several. The failure mode worth catching was two polite regions in one card (the narration's aria-live and the "Finding this step…" line's role="status", which is also polite); they never coexist once a step settles. EVE-VIS-146 (S2), fixed: with the preference set, a tour still smooth-scrolled 13,989px through 45–59 intermediate positions — the whole height of Home — for someone who had asked for less motion. globals.css already carries a universal transition-duration: 0.01ms !important; scroll-behavior: auto !important under that media query, and that is what flattens the spotlight ring (measured identical with and without a conditional in the component, so the conditional I first wrote was removed rather than kept as decoration). But a CSS scroll-behavior cannot override a behavior argument passed to scrollIntoView, so the one motion the tour drives from JavaScript was the one that escaped. The harness was measuring nothing, which is why this was nearly missed: test.use({ reducedMotion }) does not reach the page in this config — the context option leaves matchMedia('(prefers-reduced-motion: reduce)').matches FALSE while page.emulateMedia() sets it true — so the entire "reduce" arm ran at the default, and its one failing assertion looked like a product defect a weaker assertion would have "fixed". The cell now emulates explicitly and asserts the preference is in force before measuring: a preference cell that cannot prove the preference is set is measuring nothing. That is the fourth measurement-wrong-in-the-passing-direction of this session and the worst of them, because the arm it silenced was the accessibility one. Both arms green; scroll positions 42–59 before the fix, 2 after; calibrated by restoring behavior: 'smooth'.)
  • 5.5 TTS narration path (voice-enabled session): narration plays per step, stops on End/Esc, never overlaps itself. (CLOSED 2026-08-09 s20. One of the three claims held — the per-step guard from 5.1 means a step narrates exactly once however many times it re-locates. The other two were both false. EVE-VIS-147 (S2), fixed: advancing a three-step tour with voice on played three narrations at once, and Escape left all three still playing over a screen with no tour on it. speakViaBrowser cancels the synthesiser first, but the SERVER path — the default whenever there is a real session — built a new Audio, overwrote the ref and played it, leaving the previous element sounding and unreferenced. Stopping the CURRENT audio is not enough, and a first fix that only did that changed the tape not at all: speech is FETCHED before it is played, so when a new request starts the previous one has no element to stop yet. Each request now claims a generation (the device responseGenerationRef already uses for superseded turns), a stale response is dropped rather than played, and stopSpeaking retires everything in flight as well as what is sounding; handleTourExit stops speech. Both of this task's findings were nearly lost to the instrument, in opposite directions. The tape recorded pause but not ended, so three SEQUENTIAL narrations read as three concurrent ones — the first "defect" it reported was its own arithmetic, and it survived a fix that did nothing, which is what exposed it. And the doubled TTS body was a 44-byte WAV that ends the instant it starts, which made overlap physically impossible and let the "stops on End/Esc" assertion pass against code that had no mechanism to stop anything — it was measuring the clip's length. With ended taped and a ten-second clip (about as long as narration), the pre-fix numbers are unambiguous: peak 3, three still playing after Escape; after: peak 1, none. The exit leg now states its precondition and fails loudly if nothing was sounding when Escape was pressed, rather than banking the pass.)
  • 5.6 Member-authored tours: submit via conversation → operator review queue UI (admin) → reject-with-note copy → approve → appears in list_curated_tours → runs with revalidation; a stale-anchor approved tour refuses at start with honest copy. (CLOSED 2026-08-09 s21. The arrow in this task's own title has a gap in the middle of it: the operator review queue UI did not exist. Both endpoints have been live and admin-scoped since P2 and no admin page called either, so the only way to review a member's tour was curl — EVE-VIS-148 (S2) — and every submission sat pending forever while the assistant told its author, truthfully, that a human had to approve it. Built: a queue on /admin-tools with two same-origin proxies, printing every step's narration and the anchor it points at (a queue that shows "3 steps" and an Approve button is a rubber stamp), requiring the note the BFF requires before the request is sent, and running start_tour's OWN validator over already-approved rows so a tour that has gone stale is seen by an operator rather than met by a member. The member's side had two more holes, both of them halves of loops. EVE-VIS-149 (S2): nothing anywhere read a submission back to its author — listByAuthor had no production consumer at all — so the note the store REQUIRES on a rejection, and whose comment calls it "a note the author can act on", reached nobody. list_my_tour_submissions closes it, and the operator's own sentence is now asserted on the member's screen in all six cells. EVE-VIS-150 (S2): list_curated_tours offered every approved member tour whether or not it still validated, and start_tour refused with "approved tour "tour…" no longer validates: steps[0].anchorId … is not in the anchor registry" — the offer arriving before the refusal (EVE-VIS-057's shape) in a sentence the model can only relay (EVE-VIS-129's). Plus EVE-VIS-151 (S3): two id shapes come back from one list, the model put a member tour's id in the curated field and told the member "Let me try that with the right reference" — the id spaces do not overlap, so start_tour now resolves one id whichever field it arrived in. The stale tour is real, not simulated: an approved row restored from the durable snapshot whose second step points at a nav the shell removed, which is the only way one can exist (the store validates at submission; restoreSnapshot does not, because the registry it would validate against is the one that has since changed). Six member cells green (cream and dusk × 1470/390/735, five live turns each) and two operator cells (desktop, narrow), decisions read back out of the store rather than off the page. A defect I nearly did not look for came out of the dusk cells: EVE-VIS-152 (S2) — the Send button measured 2.24:1 in all three, because the assistant surface is conditionally rendered INSIDE the panel, a tour replaces it with a new <aside>, and the effect publishing --assistant-on-accent had dependencies that do not change when the element does; its work stayed on the element that had been discarded and the glyph fell back to the theme's paper, which is dark in dusk. 2.24:1 was already written in that effect's own comment as the ratio it exists to prevent. Two admin-shell defects were found and measured rather than swept: EVE-VIS-153 (S2, fixed) — a hydration mismatch React said it would not patch up, on every admin page at 390px, from a guard reading window.innerWidth during render — and EVE-VIS-154 (S2, open, owner 9.1) — every admin page scrolls sideways 83px at 1470 from the header's right cluster, reproduced on /dashboard with none of this task's code, so the layout lens here is scoped to the panel with that scoping written into the spec. EVE-VIS-155 (S3, open): an approval cannot be taken back. The measurement that changed my own verdict. The recon transcript had the assistant promising "I'll let you know once it's live in the shared catalog" — a notification with no sender — and I was ready to ledger it as a defect the fix had closed. Run properly at n=20 per arm through the BFF's own /turns, one fresh member per sample: 0/20 before and 0/20 after. The model is usually honest and one occurrence in ~21 is below what 20 samples can see, so it is recorded as an observation and explicitly NOT as a rate. What the fix DID change is measured: 0/20 → 20/20 replies now tell the member they can ask. Three instruments were wrong before the product was, each caught by its own control: the promise detector cleared "I'll flag it for you when it's live" because it required "flag you" (its positive control failed on the very sentence the finding came from); the approval poll watched an on-screen count the stale row had already made non-zero, so it returned before the POST landed and reported an approval that was demonstrably in the store a second later; and the tour-title read took aria-label off the overlay instead of the card, reporting a tour that was on screen as "(unnamed tour)". A fourth was caught before it shipped: the first hydration lock could not fail, because typeof window === 'undefined' is FALSE inside renderToString under jsdom. Also recorded honestly: the six-cell vocabulary guard caught MY OWN new tool field — the model said "it's no longer startable", so the key is canRun; the journey-inventory freshness gate is red and was verified red on a stashed tree first (EVE-VIS-125, unchanged); and the harness's 40 pending submissions were cleaned up the way 9.6 requires, rejected with a note naming the harness rather than deleted.)_
  • 5.7 Captions/karaoke rendering (never browser-tested): with word timings available, captions render word-synced; without timings, honest fallback; legibility over busy backdrops; both themes. Close the standing "karaoke rendering pending browser pass" item. (CLOSED 2026-08-09 s21. There was nothing to browser-test: the rendering did not exist. captionWindowAt shipped on 2026-08-04 with word timings flowing end to end, and its only importer in the whole repo was its own unit spec — EVE-VIS-156, the EVE-VIS-118/121/149 family again, machinery nobody reads. Which timings can honestly be synced is the whole question, and it took reading both speech paths to settle. The product speaks two ways. speechSynthesis fires boundary as it talks, carrying the character index it has reached — that is a real measurement of where the voice is, and the caption now follows it word by word. Server speech is an Audio element and a content type: synthesize(text) returns no word timings, neither provider reports them, and estimating them from duration and word length would put the emphasis on the wrong word for anyone who can hear it and on a guess for anyone who cannot — so that path renders the reply with nothing highlighted and says which it is in the DOM (data-assistant-caption="plain" against "synced"). That is the task's "honest fallback", arrived at by refusing to fabricate an alignment rather than by failing to build one. One window implementation serves both: captionWindowAtIndex was extracted and captionWindowAt now delegates to it, so a caption placed by a clock and one placed by an index cannot drift a word apart — and the test that asserts they agree caught my own confusion between the absolute index and the window-relative one on its first run. Six cells green (cream and dusk × 1470/390/735), both arms per cell, five lenses, console clean: 12.53–15.32:1 on the caption text in every cell, the caption hit-testing to itself everywhere, and the window moving with the voice rather than pinning the opening. Doubles are at the dependency boundary and named: the TTS endpoint serves a real ten-second WAV (task 5.5's double) so the server arm genuinely plays, and the SYNTHESISER is replaced in the page because headless Chromium ships one with no voices that never fires a single boundary — the tokenising, the mapping, the window and the rendering are the product's throughout. Two of my own bugs, both caught by a measurement rather than by looking: the caption's centre hit-tested to the composer (sticky at bottom: 3.6rem with the same z-index as the bar it sat behind — it now shares the composer's sticky block), and spacing the words with margin-inline-end made the caption's own text read "Trytheten-minutesit" to anything that reads it rather than paints it. A third instrument fault was mine too: the first run measured the synced arm as "plain" because the previous arm's ten-second audio was still playing and its caption was still on screen, so the wait returned against the turn before. Recorded rather than built: the member's own recorded speech has real Deepgram word timings and no surface plays it back, so those timings still have no consumer; giving them one means the review affordance task 6.2 owns.)

Phase 6 — Voice: every state of the pipeline#

  • 6.1 Mic permission flows: first-ask, denied (honest guidance copy), revoked mid-session. (CLOSED 2026-08-09 s21. Two of the three cases were broken, and the copy this task asks about existed the whole time — it was on a branch nothing could reach. EVE-VIS-157 (S1): with the microphone blocked, startServerRecording's catch handed over to the BROWSER recognizer, which needs the same permission that was just refused. Chromium accepts start() on a blocked mic and then says nothing at all, so for the fifteen seconds of the silence watchdog the member had a button reading "Stop listening", a composer greyed out and impossible to type into, and not one word about why — and the note that finally arrived was the watchdog's, which is about quiet, not about permission. The honest sentence EVE-VIS-042 wrote for the recognizer's own not-allowed branch is now said immediately, and only a real capture failure (no device, a mic another app holds) still hands over. EVE-VIS-158 (S2): nothing watched for the microphone going away mid-recording. The VAD cannot stand in for it — an ended track produces no samples, so the detector never sees the quiet it waits for — and the panel went on recording a dead track behind a disabled composer until the member worked it out. startVoiceRecording now takes onSourceEnded; the recording is cancelled and the member is told nothing was sent. Six cells green (cream and dusk × 1470/390/735), each driving all three cases: the prompt is not claimed as listening while it is open, a revocation returns the panel to idle with an explanation, a refusal explains itself in one ask (asks === 1, so the browser is not pestered for a permission it just declined), the composer takes typing straight after, and no raw error name (NotAllowedError, getUserMedia, DOMException) reaches the member. The refusal is driven at getUserMedia deliberately: this harness runs Chromium with --use-fake-ui-for-media-stream so the recording probes work at all, and that flag auto-accepts every prompt — a permission decision cannot be exercised here in either direction without doubling the gate the browser owns. Two harness faults, both of which made cells measure the wrong thing: the first version remembered only the LAST stream getUserMedia handed out, and the panel opens the microphone twice in parallel (recorder + level meter, now EVE-VIS-159, S3 open) — so revoking "the" stream ended the waveform's and left the recorder's alone, and two cells reported the product ignoring a revocation it had never received. And tapping the mic before the assistant session exists routes to the browser recognizer instead of the server recorder, which sets "listening" with no permission round trip at all; two cells measured that and reported the panel as claiming to listen during the prompt. Both are written into the spec. Also fixed in the environment, not the product: the seeded members were free-plan, so every upload came back assistant_voice_not_in_plan and the revocation case ended in a plan notice rather than in what the microphone did — the BFF now runs with OSHUN_DEV_SIGNUP_PLAN=pro, which task 4.6 had already found necessary.)
  • 6.2 Push-to-talk UI states: idle, recording (level indication), processing, transcript-arrived (member sees the transcript BEFORE it sends as a turn — verify the review affordance), error. (CLOSED 2026-08-09 s21. Two of the five states did not exist, including the one the task text describes. EVE-VIS-161 (S2): both speech paths called sendMessage(transcript) the instant the words arrived, so the member never saw what had been heard — a misheard sentence went out as their own words with no moment in which to stop it, and a member dictating a name or working in a noisy room found out what was understood only when the reply came back about something else. The transcript now lands in the composer as a draft, focused, cursor at the end; Send is the affordance they already know and correcting a word is typing rather than starting again. Two details the first version got wrong: it must not clobber text typed before the mic was reached for (it appends), and the browser recognizer's FINAL result must REPLACE the interim guesses it has been writing into the composer all along. EVE-VIS-160 (S2): there was no processing state at all — finishServerRecording set listening false and then awaited an upload and a transcription, so for the whole round trip the member watched an idle microphone over a composer they could not type into. Three states now, in the composer's own prompt and in the screen-reader hint beside it, cleared in a finally so every outcome leaves it. EVE-VIS-162 (S2), found by this spec's own first run: a recording that captured nothing ended in a bare return — the one branch EVE-VIS-042's sweep did not reach. Six cells green (cream and dusk × 1470/390/735) walking idle → recording → processing → the transcript to check → sent on the member's tap → an error, five lenses and a clean console on all four inspected states. The level meter paints 12 real bars while recording and none after, so its absence stays honest. The STT endpoint is doubled (as 5.5 and 5.7 doubled TTS) because this deploy has no Deepgram or Whisper credentials — without it every cell would measure the 503 hand-over to browser speech instead of the states under test — and the delay on it is what makes the processing window observable. The harness taught me something the product then had to answer for: a 1.2-second recording produced no transcript, no note and no state change, which is how EVE-VIS-162 was found; the recording wait is now three seconds, with the reason written down (MediaRecorder is given no timeslice, so it flushes on stop and a very short take can come back empty). And one case is deliberately locked in jsdom rather than in the browser: the empty recording could not be reproduced on demand — 250ms still produced audio — and a browser case that never fails is a case that passes against the broken build.)
  • 6.3 Server STT path (Deepgram/Whisper binding) and 503 → Web Speech fallback with its honest system notice — both visually verified. (CLOSED 2026-08-09 s22. The fallback's notice described the machinery and hid the consequence, and the path it hands over to was announced even where it does not exist. EVE-VIS-163 (S2): "switching to your browser's speech recognition" is read as "it is being handled another way", and it is not — the refused recording is gone, the recognizer now opening heard none of it, and the microphone is live in front of a member who has been given no reason to speak again. The notice now says the words were not heard, that the mic is open, and — because the composer is disabled for as long as the recognizer holds it — names the step that frees the keyboard. EVE-VIS-164 (S2): neither hand-over branch asked whether the browser HAS a recognizer, so on Firefox (constructor behind a preference, off by default) every refused recording produced a notice that the next line retracted. EVE-VIS-165 (S2), found by a failing cell and then measured frame by frame: the transcript draft was focused on one requestAnimationFrame and the composer is disabled while listening — focus() on a disabled element does nothing, silently, and over three identical runs it worked twice and failed once. That is task 6.2's own affordance, passing by luck. EVE-VIS-166 (S3): five-line notices centre-aligned. EVE-VIS-167 (S2), found by running the server leg against the real vendor: a 502 answered "please try again" every time, with no memory and no fallback — the first is fair, the second now hands over. The server-STT leg is real up to the vendor and no further, deliberately. Chromium's microphone plays a real WAV of the sentence (--use-file-for-fake-audio-capture; the silence either side is kept under the 1.8s VAD hangover so the probe, not the detector, ends the recording); the panel records it, uploads ~110 KB of webm to the real route, and the real WhisperProvider posts a real multipart request — which the double ASSERTS rather than merely answers (audio.webm, audio/webm, model=whisper-1, response_format=verbose_json, language=en, >10 KB), because a route-level double like 6.2's cannot see the binding at all. Only OpenAI is replaced, via a new OSHUN_ASSISTANT_STT_ENDPOINT (a real deploy knob — Azure OpenAI, Groq, self-hosted whisper.cpp, Deepgram on-prem — that both providers already accepted and nothing passed). Why: the live account answered "You have no credits remaining", which is the vendor's state, not the product's; it is what found EVE-VIS-167, and a probe whose red depends on an unpaid invoice is not a probe. Deepgram is unexercised on this box (no key); its endpoint wiring is unit-tested. Sixteen live cells green: six server STT + six hand-over (cream/dusk × 1470/390/735), four provider-failure, plus a no-recognizer browser and the recognizer Chromium actually has — which accepts start() and then says nothing at all, leaving the member 15.0s with a disabled composer before the watchdog speaks. That recovery is honest and bounded, so it is recorded rather than ledgered. Two things I got wrong and the instruments corrected: a cell sampled the processing state at a guessed 700ms, read "Ask Lilith anything" and briefly looked like a missing state — measured frame by frame the panel enters it 200ms after the stop and the local round trip was simply over, so the double now takes a realistic 1.5s and the state is waited for by name; and the composer's ::placeholder, which carries the ONLY words in the listening and processing states and which the five-lens contrast measurement never reads (it reads color, i.e. typed text), was measured directly and passes comfortably — 7.34:1 cream, 7.00:1 dusk.)
  • 6.4 Silence auto-stop: speak then stop — auto-stop at ~1.8s quiet; opt-in surfaced correctly; no premature cutoffs on natural pauses (test 3 cadences). (CLOSED 2026-08-10 s22. The detector was never the problem — its stop went nowhere. EVE-VIS-168 (S1): the VAD fired on time, tore down its analyser and stopped the MediaRecorder, and nothing told the panel. The mic went on reading "Stop listening" over a composer that could not be typed into, and the level meter kept painting bars from its own separate capture — so the member had positive visual confirmation of being heard while everything they said was recorded by nothing, and not a byte was uploaded until they gave up and pressed stop by hand. The recorder had one ending its caller knew about and two it did not (the hangover, and the one-minute cap); onAutoStop(reason) is now called by both. EVE-VIS-169 (S3): the behaviour this task calls an opt-in was never surfaced — "Listening…" said nothing about the microphone closing itself, so a member reads their own pause as an interruption and hurries, or waits for a stop they never needed. Three cadences, built from the same two half-sentences so only the gap differs, played into Chromium's microphone as real WAVs and never touched again after the first tap: a clean stop, a 0.6s breath, and the 1.4s pause a member takes to think. All three now close on their own, none inside its gap, each within half a second of its due time (A 5367ms against 4880ms due; B 5778/5480; C 6497/6280 — the overshoot is the detector's 100ms sampling plus the recorder's flush), each uploading exactly once and each ending in a draft to check. Six cells, cream and dusk. The instrument was calibrated before any of this was called a defect, because the first run was equally consistent with a broken product and a fake microphone that was not playing the file: --use-file-for-fake-audio-capture really does play it (~2.7s above the 0.045 speech threshold, then true digital silence), where Chromium's default source is a beep that never goes quiet and could not have exercised a silence detector at all. What finally pinned it was tagging every AudioContext with its creation stack and timing its state changes: the VAD's own context read 5ms:running, 4850ms:closed on a cadence whose speech ends at 3080ms — the 1.8s hangover to within 30ms.)
  • 6.5 TTS playback: speaker states in the panel, interrupt-on-new- turn, plan-gated 403 copy for free plan (assistant_voice_not_in_ plan must render as friendly upgrade copy, not an error code). (CLOSED 2026-08-10 s23. Three defects found and fixed in s22; the question s22 refused to guess at is now measured, and the answer is that the panel was never the problem. EVE-VIS-170 (S2): the streaming turn route sent shouldSpeak: inputMode === 'voice', so a member who turned voice output on in the header and then TYPED got nothing read aloud — their own toggle overruled by how they had asked — while the deterministic route sent the formatter's true for the same turn, making it depend on whether streaming happened to work. Found by the first run of this spec waiting 120s for a speaking state that could not exist: the turn had completed and no TTS request was made at all. EVE-VIS-171 (S2): speakViaServer's catch handled 503 and nothing else, so a free plan's 403 fell through to the generic branch — every reply paid for a doomed round trip, the voice quietly became the browser's instead of hers, and nothing ever said why. The STT side has had that branch since EVE-VIS-054. EVE-VIS-172 (S2): sending a new message did not stop the previous reply from speaking, so the member asked something else and was read the answer to the question they had moved on from; now stopped for messages the MEMBER wrote, and only those, because a tour re-plan goes through the same function while the tour is narrating. Every cell has passed (cream and dusk × 1470/390/735) and each leg is mutation-calibrated. The composer wait: measured, not guessed. s22 left this open because a turn still running and a sending flag that never cleared produce an identical screenshot, and one of them is a member who cannot type in front of a finished reply. They are told apart by putting both ends of the wait on the page's ONE clock — the wire observer timestamps turn.complete as its bytes are parsed, and a new 100ms panel sampler (support/panel-timeline.ts) timestamps the controls. With text in the box, disabled={!input.trim() || sending} can only mean sending, so the interval from typing to a usable Send IS the flag. Across 24 waits in six cells: where nothing was in flight, Send freed in 4–19ms; where a turn was still running, the composer came back 200–476ms after its terminal frame. The panel never held the composer for nothing. The long waits are the client's own budget working — 30.6s, 33.1s, 33.2s and 43.2s are the stream client's 30s idle abort plus the deterministic route's hand-over, and their cause was environmental (getaddrinfo ENOTFOUND openrouter.ai, intermittent DNS on this box, 3 of 4 turns in one cell). Throughout every one of them the member had the typing indicator AND the interrupt control (interruptOffered true on all 11 waits over 3s), so a long wait is visible and escapable rather than a dead composer. s22's 240s was the harness typing too early — the fault it had already found and fixed in the same session — and this run, on the fixed harness, never came within 200s of it. Two instrument faults were caught before either could be reported as a defect, which is why the first six-cell run of this analysis went red in 4 cells: "a reply had already arrived" was true from turn 1 onward whatever the CURRENT turn was doing, and the wire observer watched only /turns, so the four seconds a failed stream spends in the deterministic /message hand-over read as four seconds of nothing. Both are fixed in the support modules; in-flight is now counted across BOTH routes. The property is now a lock, and the lock is calibrated: delaying the finally's setSending(false) by 20s turns it red with "no turn was in flight and Send stayed disabled for 18638ms".)
  • 6.6 60s cap behavior and the 10MB/415 guardrails surface humanely. (CLOSED 2026-08-10 s23. All three surfaced as the same sentence, and it was the wrong one for every one of them. EVE-VIS-173 (S2): the recorder hard-caps at 60s and reports which ending it took — onAutoStop('max-duration') — and the panel's handler took no argument, so the cap was indistinguishable from the silence auto-stop. The microphone simply shut, mid-word, and the transcript that arrived just stopped; a member dictating a passage had no way to know there is a minute, that they had met it, or whether the rest was kept — while the composer said "I stop when you pause", which describes the other ending. EVE-VIS-174 (S2): a 413 from the 10 MB limit fell through to "That recording could not be transcribed … Please try again", which is an instruction to repeat a failure that will be exactly the same size next time, and it counted toward the provider-failure tally, so two of them would have retired server transcription over something the server never looked at. EVE-VIS-175 (S2): a 415 got the same sentence, and that refusal belongs to the BROWSER — every attempt from here produces the same container — so it was a wall with no door in it, the exact shape EVE-VIS-167 was opened for on the branch its fix did not reach; it now hands over as 503 and 403 do. The cap is driven for real in all six cells (cream/dusk × 1470/390/735): Chromium's DEFAULT fake microphone is a tone that never goes quiet — the very property that forced 6.4's probes to play a FILE instead — which makes it the one source on which the VAD cannot fire, so a recording that ends at 60s ended because the cap ended it. Measured 59,951–60,011ms against a 60,000ms cap, one upload each, 957–958 KB of real opus reaching the real route and the doubled vendor. 413 and 415 are doubled at the route, and the honest reason is that neither can be reached from this client at all — that same capped minute is ~958 KB against a 10 MB limit, and every container MediaRecorder produces here is on the server's accept list — which is exactly why nobody had looked at them. The server's own half is now held too: the 10 MB limit had no test, and a bodyLimit is a number in an options object that deletes without breaking anything. Adding one found that it is declared TWICE at the same threshold, on the content-type parser and on the route, so each masks the other — two separate controls came back green while proving nothing, and only raising BOTH to 64 MB turns the test red. A sweep alongside puts the boundary at exactly 10,485,760 (9,000,000 → 200, 10,485,777 → 413), so nothing else is quietly standing in front of it. Both refusals are also locked in jsdom, where they run as a CI gate rather than only in the harness, and both locks are calibrated by disabling their branch. One instrument fault, caught by looking at what passed: the 415 leg's first version accepted any of /browser|record|format|type/, and "That recording could not be transcribed" contains "record" — so the leg passed against the very copy it exists to reject. Tightened to /format/ plus an explicit refusal of the generic sentence.)

Phase 7 — Audit mode: the walk-everything experience, deeply#

  • 7.1 Full audit conversation arc across ≥3 domains and ≥10 flows: begin → status → walk → mark visited → skip-with-note → resume after panel close → resume after PAGE RELOAD (durability visible) — coverage numbers restated every turn per conduct rules. _(CLOSED 2026-08-11 s24 — green begin to reload in one run. phase-7-1-audit-arc.spec.ts on support/audit-conversation.ts. The run: 37 turns, 0 retries, 13 buffered fallbacks; 11 visited + 1 skipped over shell/tara/nyx against a floor of 10 flows and 3 domains; run id and both counts IDENTICAL across panel-close, re-open and page RELOAD; no member message moved the run by more than one flow; zero machine vocabulary; zero claim-corrections needed; console clean. Four defects closed on the way: EVE-VIS-177 (a turn that died upstream was answered with "Could you rephrase that?", blaming the member — the fallback engine now reports whether it understood at all, and the first cut of that fix shipped a claim it did not keep, caught by a route test); EVE-VIS-178 (S1 — "End tour" started another tour); EVE-VIS-179; EVE-VIS-181 (one Escape dismissed two layers, found in TWO components). EVE-VIS-183 is open with the catHeaderBar half fixed. EVE-VIS-182 was withdrawn on the evidence rather than filed.

    **Five harness faults, every one of which would have been reported as a
    product defect.** (1) Coverage reads held a seed-time access token that
    expires ~15 minutes into a 20-minute arc: run 4 watched `visited` climb to
    6 and then read `null` for every turn from 23 on, while the panel beside
    it went on calling `audit_status` happily — the run was fine and the
    harness's reading of it had stopped working. It now refreshes through the
    app's own `/api/auth/session`, and only on a 401, so a dead session still
    looks dead. (2) A turn was re-asked whenever it called no tool — which is
    equally true of a turn that ANSWERED and asked the member a question. Run
    4 sent the same sentence three times before the agent said "I can't just
    keep cycling on the same line". EVE-VIS-177's own fix supplied the exact
    signal to tell the two apart. (3) `closePanel` waited for the composer to
    vanish, which a running TOUR also does, so it passed on a panel that never
    closed. (4) **The walk did not walk.** It answered every turn with "that
    one works too, mark it visited", the catalog's first flow has five steps,
    and the agent refused all forty times — correctly, and in as many words:
    "I won't rubber-stamp it", "are you on the Library tab? plain yes or no".
    Run 7 finished 0 visited. The harness now performs the flow's steps and
    names them; the product was never the thing that was wrong. (5) The
    off-screen lens flagged the opening chat bubble at y = -28, i.e. the
    scroller working — it is a CONTROL lens and a transcript is not a control.
    
    **Two things the product does that are right and were nearly filed as
    wrong**, both settled by looking rather than by reasoning from stacking
    numbers: a fullscreen room overlay covering the docked assistant
    (EVE-VIS-182, withdrawn), and Tara's session player declining to end a
    meditation on a stray Escape. **Landmine:** `e2e-inspect/.evidence/` is
    gitignored and disappeared mid-session — cause unconfirmed, but every
    ledger row citing a path in there should be assumed regenerable, not
    durable.)_
    
  • 7.2 /assistant/audit board: every element five-lens inspected — progress bars, per-domain rollups (all rows, not just the top), next-pending card with steps/routes/anchors, run metadata line, no-run empty state, begin CTA, chat handoff BOTH directions ("Start the audit" → panel opens primed; "Continue this flow in chat" → correct resume prompt). _(CLOSED 2026-08-11 s25 — 12 cells green. phase-7-2-audit-board.spec.ts: the board in cream/dusk × 1470/390/735, the no-run state in two, the handoff, and the completed state. Read against a REAL run — 755 flows over 162 domains, 2 visited and 3 skipped, the visits genuinely walked and the skip reason the member's own — with every rendered number compared to the coverage payload: 163 bars, and for each of the 162 domain rows the label, the N done · N skipped · N pending line, aria-valuemax and aria-valuenow.

    **Four defects closed.** **EVE-VIS-184 (S1)**: every hard load of this
    route showed a signed-in member "Audit coverage is unavailable:
    Authentication required". The board fetched on mount and the auth
    context's token mirror is empty until `/api/auth/session` returns — the
    board won that race 3 of 3 loads, and the Try again button worked every
    time, which is the shape of a race rather than a failure. **EVE-VIS-185
    (S2)**: the empty state promised "164 rooms · 760 flows" and the button
    built 755 across 162 — the difference is the five Veritas/Metis flows
    V1.2 defers, which the server has reported as `deferred` since
    EVE-VIS-176 and the web client's type did not even carry. Both states now
    say it. The same line cost **4,179 KiB** (the entire catalog graph,
    reduced client-side to four integers) and now costs **116 bytes** through
    a new `GET /v1/assistant/audit-runs/scale` derived from the very set
    `beginRun` materialises. **EVE-VIS-186 (S2)**: the next-pending card is
    the instruction sheet for a flow the agent will not mark unless it was
    really walked, and it rendered step TITLES only — dropping the `action`
    ("Go to the Home tab and observe the daily rail") and the `anchorId`. The
    anchor is now shown as the shared registry's own sentence about that
    control, never the raw id, and the route is a link instead of inert mono
    text. **EVE-VIS-187 (S3)**: ninety-odd rows read "1 suites"; fixed in the
    e2e compiler and the product-graph artifact regenerated (gate green).
    
    **One defect withdrawn on the evidence.** Four of six cells reported the
    board's own "Continue this flow in chat" button buried under the fixed
    mobile nav. It is not: `probe-7-2-cta-under-nav.spec.ts` clicked it for
    real in all four and the panel opened. The card is the last thing on a
    9,800px document, so `scrollIntoViewIfNeeded` lands its footer under the
    nav and one more notch frees it. **Covered is not the question;
    reachable is** — the matrix now judges controls with a TRIAL click, which
    is Playwright's own actionability loop rather than one hit-test at
    whatever offset the inspection happened to stop at.
    
    **Two more harness faults, both in the direction of false defects.** (1)
    The off-screen lens is a CONTROL lens and this page is a document: on a
    163-bar board it flags every row below the fold. Only sideways escape is
    a finding here. (2) A coverage bar is two flex SIBLINGS, and at 390px a
    run of 2-in-755 gives the first one a width of 0.9 CSS px —
    `elementFromPoint` at its centre rounds onto its neighbour and reported
    the bar as painted over itself in all four narrow cells. Sub-pixel
    slivers cannot be buried; their widths are asserted from the ARIA values.
    
    **Seeding the board cost three runs and taught one thing worth keeping:**
    report the walk from the SHELL, not from the page you finished on. Run 2
    said "I have just been through 'Search across domains' — I opened Open
    search, Run a query…" while standing on `/search`, and the agent answered
    "Found 2 results for 'across domains'" three turns running. It read the
    report as a query, and on that page that is a reasonable thing to do.
    From home it marked two flows in two turns. `walkFlowSteps` also now
    names what blocked a step and falls back to direct navigation — its first
    failure was a tour whose spotlight takes pointer events for the whole
    viewport, and the walk gave up on all five steps rather than saying so.
    
    **A completed run is 755 marked flows away**, so the completion state is
    driven by fulfilling `/audit-runs/current` with a payload derived FROM
    the real one — real domain titles, real totals, outcomes redistributed.
    The dependency is doubled; the board under test is not.
    
    **Open, and not this task's:** EVE-VIS-188 — `compilers.spec.ts`'s
    TOTALITY gate is red on a clean tree (`expected 1108 to be 1107`),
    confirmed to predate this session with `git stash push -u`. Its constant
    is deliberately written down so estate growth is visible, so the repair
    is to account for the new verify edge, not to bump the number.)_
    
  • 7.3 Drift rendering (post-fix): the warning appears ONLY on a genuinely drifted run; the E4 drifted data (count/flowIds) surfaces intelligibly on the board and in chat; a drifted-visited flow never re-walks. _(CLOSED 2026-08-11 s25. phase-7-3-drift.spec.ts (6 cells) plus locks in audit-store.spec.ts, audit-agent-tools.spec.ts, AuditCoverageBoard.spec.tsx and audit-coverage.spec.ts.

    **Three things get called drift and they are not the same thing.** The
    VERSION note says the catalog moved; DRIFT says flows in THIS run are
    gone from it; and the run's own pin is a third. The board could only make
    the first claim — `drifted` was not in the web client's type at all, so
    the data the server has published since EVE-VIS-176 reached nothing. It
    is now a sentence with the count in it, the ids behind a "Which ones?"
    disclosure (they are the only identity a vanished flow has left, and they
    are builder vocabulary — they do not belong in the prose), and an honest
    "Showing the first 25 of N" when the server's cap bites. The undrifted
    run asserts the negative in both themes: no version note, no drift
    sentence, no disclosure, and the words "removed from the catalog" nowhere
    on the board.
    
    **The chat half was a branch that had never been exercised.**
    `compactCoverageForTool` attaches `drifted` and a `driftNote` telling the
    model to say so rather than re-walk, and nothing asserted either — a note
    the model is never given is indistinguishable from one it ignores. Now
    locked, and calibrated by deleting the branch. The same test found the
    cap unstated: 40 gone, 25 named, and a model handed that will say "these
    are the ones". The note now says which it is.
    
    **EVE-VIS-189 (S1), found by walking rather than by looking.** A run
    completes when nothing is pending; a pending item whose flow has left the
    catalog can NEVER be marked — `nextPending` steps over it, correctly, and
    `markItem` is never called for it. `beginRun` resumes an active run. So
    the member could neither finish the audit nor start another, for ever,
    and the board's advice at that point was "open the assistant and ask for
    the audit status" — a conversation with exactly the same nothing to
    offer. Closed with a real exit: `closeRun` + `POST
    /v1/assistant/audit-runs/current/close` + "Close this run and start a new
    one", which closes BEFORE beginning (the other order resumes the stuck
    run — asserted on invocation order). Closing is not completion: the run
    keeps its unresolved items and the board says "closed with N unresolved
    flows" instead of "every flow visited or skipped". Reachable today
    without any catalog edit, by any run begun before this release's room
    cut.
    
    **EVE-VIS-190, withdrawn on the evidence, and it cost an hour.** The
    dead-end screenshot showed the member's own opening line twice, and the
    network agreed: two turn POSTs to TWO DIFFERENT session ids. Measured
    both ways against the same build — `reactStrictMode: true` → 2/2/2,
    `false` → 1/1/1. It is React's development double-invoke, which does not
    happen in a production build. What doubles is the panel's session
    creation, so **any dev-mode measurement of turn counts is 2×**;
    `probe-7-3-primed-duplicate.spec.ts` records it so nobody pays for it
    again.
    
    A drifted run cannot be produced against a running server without editing
    its sources, so the drifted and dead-end BOARDS are presented by
    fulfilling `GET /audit-runs/current` from the member's real coverage —
    and the exit out of the dead end is then driven for real, double removed,
    against the real BFF.)_
    
  • 7.4 Long-run ergonomics: with 50+ marked flows, board render performance, list virtualization/scroll, and the chat's coverage summaries stay readable (top-10 domain cap note reads well). _(CLOSED 2026-08-11 s25. phase-7-4-long-run.spec.ts (5 cells) plus probe-7-4-history-cost.spec.ts and probe-7-4-refused-turn.spec.ts.

    **Virtualization is not needed, and now there is a number saying so.**
    159 bars, 1,388–1,480 DOM nodes, an ~8,900px document: a full-document
    scroll of 40 frames costs **658 ms — 16.5 ms a frame, with zero long
    tasks**, in both cream/1470 and dusk/390. The budget is asserted (4 s for
    the scroll, 350 ms for any single main-thread block), so if the list ever
    does need virtualizing, this is the test that will say it.
    
    **The summaries are readable and true**, checked against a plumbing
    detector with its own positive controls — because the cap note the tool
    attaches ends "…the full rollup lives at GET
    /v1/assistant/audit-runs/current", an HTTP endpoint sitting one line from
    the numbers the model is about to read out. Nothing leaked in either
    turn, every stated count matched the run, and the reply stayed a summary
    rather than becoming a 158-row wall.
    
    **EVE-VIS-191 (S2), and the model found it.** Asked which rooms had the
    most left, the agent answered "Veritas — 20 (a room that isn't in this
    version yet, so it can wait)" — true, and about work that should never
    have been in the run. EVE-VIS-176's fix asks whether a DOMAIN ID is
    deferred, an exact match against `['veritas','metis']`, and the composed
    catalog's e2e half names its domains `e2e-customer-web.veritas`. So 34
    journeys about the deferred rooms sat in every run, their routes
    answering 307 → `/release-scope/…`. A room is deferred by its ROUTES as
    much as by its id, and only the web app could ask that question:
    `isRouteDeferredFromV1` now lives in `@oshun/navigation` and the web
    resolver delegates to it. **755 flows / 162 rooms → 721 / 158**, with the
    difference stated as deferred rather than subtracted in silence. The
    route test alone caught 4 of the 34 — most e2e journeys mine no room
    route — so the group is tested too.
    
    **EVE-VIS-192 (S2), found by the fix being slow.** `getCoverage`
    re-derived the release cut on every call: two `O(flows × graph)` walks
    over a 7,700-node graph, **1,042 ms per call on the code as it shipped**,
    paid by every `audit_status`, every `audit_mark` and every read of the
    coverage endpoint. It is synchronous CPU on the event loop, so the cost
    is head-of-line blocking for everything else in flight, the SSE stream
    included. The cut is a pure function of a graph that never changes:
    cached, **0.8 ms**, with a 100 ms budget locked in.
    
    **EVE-VIS-193, withdrawn — and it is the reason to keep measuring.** The
    summary leg timed out at 180 s three times against the session's
    most-used member and finished in 32 s against a fresh one, so the
    hypothesis wrote itself: long history, slow turns. Timed properly, the
    "veteran" had **seven messages** and its turns were being refused **403
    in 3 ms** by the abuse layer, on the streaming route and again on the
    buffered fallback. Not slow. Refused. The window resets, so the panel's
    behaviour under refusal stays honestly unmeasured rather than assumed.
    Two harness lessons: a spec that waits only for a terminal SSE frame
    cannot tell a refused turn from a slow one, and a conversational leg that
    shares a member with a whole suite is measuring the suite.
    
    The 50+-marks board is presented from the member's real coverage with the
    outcomes redistributed — 60 marks over 20 rooms — because 50 marks are 50
    turns away; every rendered row is still compared with that payload.)_
    
  • 7.5 Skip-note quality loop: the model asks for a reason, renders it back, and the board shows the note on the skipped row. _(CLOSED 2026-08-11 s25 — the loop closed in one live run. phase-7-5-skip-notes.spec.ts, on a fresh member and a fresh run:

    > "Skip the next flow." → **"Happy to skip it — what's your reason for
    > passing on this one?"** (nothing recorded) → "Because my cat is asleep
    > on the keyboard and I refuse to move her." → **"Done — recorded with
    > your reason. That's 1 skipped, 0 visited, 720 to go. A very legitimate
    > excuse; the keyboard can wait."** → the board: **Visit every primary
    > tab — "My cat is asleep on the keyboard and I refuse to move her."**
    
    The reason is deliberately unparaphrasable. A member's own words are the
    thing being checked, and an odd sentence makes "recorded what I said"
    and "recorded something plausible" tell themselves apart on sight.
    
    **EVE-VIS-195 (S2) — the third claim was false, and in an instructive
    way.** A skip requires a note, and the note must be the member's own
    words: two carefully built guards (the store's, and
    `audit-skip-reason.ts`'s from EVE-VIS-128) writing to a row nothing
    downstream could read. `AssistantAuditCoverage` carried counts only, so
    the board said "3 skipped" and not one word about why, and a member
    returning to their own audit could not see what they had told it. The
    coverage now carries `skippedFlows` — newest first, title resolved from
    the graph because a member reads titles and ids are for tools, capped at
    50 while `skipped` stays the true count so the board can say when it is
    showing a subset.
    
    Locked at each layer and calibrated: the store spec drives the cap on
    the COMPOSED catalog, because the curated one holds ~30 flows and a test
    that cannot reach a bound of 50 is not testing it. Two more cells read
    the run Phase 7 actually built (cream/390, dusk/1470) and compare every
    rendered row against the run's own notes — and their evidence frames now
    scroll to the section first, since `inspect` shoots the viewport and on a
    159-row board the subject is a long way down.
    
    **One 7.2 lock corrected on the way:** its finished-run double built
    `total` from the real run's item count while building the rollups from
    the domains, and after 7.4's release cut those are different numbers —
    an older run holds items for deferred flows that never appear in a
    rollup and can never be marked. `visited + skipped` cannot reach that
    total; that impossibility is EVE-VIS-189. The double now portrays a run
    begun after the cut, which is coherent.)_
    

Phase 8 — Proactive help, SelectionAsk, invocation points#

  • 8.1 Fire each proactive trigger FOR REAL: rage-click (4 clicks /1.2s/24px), nav-thrash (A→B→A→B in 12s), error-dwell (visible error ≥8s) — chip renders quietly, correct grounded prompt on accept, "Not now" dismisses, "Don't offer again" persists across reload; cooldowns and the 2-per-load cap verified; suppressed while panel open; data-assistant-private areas ignored. _(CLOSED 2026-08-11 s25 — 8 cells green, no product defects found. phase-8-1-proactive-triggers.spec.ts. All three detectors fired from real gestures in a real browser: four clustered clicks (jittered inside the 24px radius, because a detector that needs identical coordinates would be useless against a hand), an A→B→A→B bounce, and a planted role="alert" left untouched past 8s. The chip renders quietly — panels = 0 in every cell — accepting sends a grounded prompt that names the struggle, "Not now" dismisses, and "Don't offer again" writes oshun.assistant.proactive-optout and survives a reload.

    **nav_thrash is clicked, never `goto`'d.** The engine's route history
    lives on an instance in a `useRef`, so a full page load destroys it: nav
    thrash is by construction a property of a SPA journey, and a cell that
    navigated with `page.goto` would be asserting a detector fires from a
    history it cannot have.
    
    **Every negative got a control**, because "no chip" is the expected
    result of five different things and of the feature not existing. The
    private-subtree cell removes the marker and repeats the identical
    gesture at the identical point — chip. The restraint cell reloads and
    repeats — chip. Without those, both would pass on a page where nothing
    is watching anything.
    
    **What the browser deliberately does NOT claim:** that the second
    hammering was refused by the CAP rather than the cooldown. Two rules
    refuse it and a browser cannot tell them apart; the engine's own spec
    separates them. Overclaiming there would have been the failure this file
    exists to avoid.
    
    **One harness fault, measured before it became a defect report.** The
    first run reported the rage detector dead. `probe-8-1-first-load.spec.ts`
    fired the same burst at three delays after hydration, on a first
    navigation and on a reload: exactly ONE of six cells came back empty —
    0 ms after `waitForPageHydrated` on a FIRST navigation. Route churn ruled
    out in the same probe (zero pushState/replaceState). `waitForPageHydrated`
    returns before this component's effect registers its listener, and nobody
    can hammer a control before the page is interactive. A 2 s settle, and
    the reason written down.)_
    
  • 8.2 Chip visual polish: position vs page chrome at all viewports, both themes, entrance animation under reduced motion. _(CLOSED 2026-08-11 s26 — one S1 defect found and fixed, EVE-VIS-196. phase-8-2-chip-visual-polish.spec.ts, 12 cells green.

    **The offer was standing on the launcher it was offering.** The chip is
    `position: fixed` at a hard-coded `right: 24 / bottom: 96`, `z-index:
    60`. At phone width the shell owns the bottom **192px** of the viewport
    and puts the utility band — the one that carries **Ask Lilith** — at
    `bottom: calc(safe-area + 104px)`, `z-index: 11`. The chip landed on it
    and covered **all six** of its controls. Not "looked covered": a
    Playwright trial click on every one of the six timed out. `ShellLayout`
    publishes `--shell-bottom-chrome` and `--shell-assistant-inset` for
    exactly this, after the PWA toast did the same thing to the same control
    (EVE-VIS-026) and room screens buried the dock (EVE-VIS-112). The chip
    read neither, so it re-committed both — at desktop it also landed inside
    the dock's column, over the persistent-context card when collapsed and
    over the open transcript when expanded.
    
    **Suppression was listening for an event the shell never sends.** The
    chip stood down on `OSHUN_ASSISTANT_OPEN_EVENT`, which only fires for
    openings routed through `openAssistantInvocation`. The shell's own "Ask
    Lilith" is `openAssistantFromPoint` setting `assistantDockMode` as React
    state, and a dock restored from storage (EVE-VIS-106) dispatches nothing
    at all — so a returning member with the panel open got offered "Ask
    Lilith" on top of Lilith, and opening the assistant from an offer left
    the offer sitting on the conversation. The panel's presence in the DOM is
    the fact; the event was only one way to hear about it.
    
    **What the position now says:** `bottom: calc(max(var(--shell-bottom-
    chrome, 0px), 72px) + 24px)` and `right: calc(var(--shell-assistant-
    inset, 0px) + 24px)`. The 72px floor is the OTHER launcher —
    `AssistantHost`'s 52px button at `bottom: 20` on shell-less pages — so
    the historical 96px survives exactly where it was right and yields
    everywhere the shell reserves something.
    
    **The entrance existed only after this task.** Measured first: the chip
    had `animation: none` and appeared at full opacity in one frame, so
    "entrance animation under reduced motion" had nothing to verify. It now
    uses the app's own `oshun-fade-in-up 200ms` at the `entrance` easing
    token, and both motion modes are measured frame by frame rather than
    asserted from the CSS: full motion ramps `opacity 0 → 1` with
    `translateY 12px → 0` over ~200ms; under `prefers-reduced-motion:
    reduce` the computed duration collapses to **1e-06s** (globals.css's
    blanket `!important` beats the inline duration) and the FIRST sampled
    frame is already opacity 1, transform identity.
    
    **The harness is calibrated, not merely green.** One cell pins the chip
    back to `right: 24 / bottom: 96` with `addStyleTag` and REQUIRES the same
    measurement to come back red — it reports all six blocked controls and
    names the assistant trigger. Without it, six green cells would be
    consistent with a sweep that measures nothing. The three unit locks were
    calibrated the same way, by mutating the built component: each mutation
    reds exactly one test, so no guard is masking another.
    
    **Two things deliberately NOT claimed.** The sweep judges FIXED surfaces
    only — page content scrolls out from under a floating offer and chrome
    cannot — and at 200% zoom the chip's top edge still tucks a few pixels
    under the header's frame, recorded rather than patched: 450 CSS px of
    height minus 192 reserved below and ~63 of header leaves 195 for a
    172.6px chip, and the only way to buy that gap is to push the offer back
    into the band this task exists to clear. Every header control there
    takes a click, measured.)_
    
  • 8.3 SelectionAsk: select text on reader/domain/settings pages — chip placement near selection, never off-screen, prefilled ask correct, private-subtree exclusion; keyboard/touch selection at narrow viewport. _(CLOSED 2026-08-12 s27 — one S2 found and fixed (EVE-VIS-197), one S2 found in passing and ledgered OPEN (EVE-VIS-198). phase-8-3-selection-ask.spec.ts, 19 cells green across all three page classes the task names — reader (/nisaba/scholar), domain (/nyx), settings (/settings) and /library — both themes, desktop and 390.

    **The offer sat inside the bar it was offering from.** The chip placed
    itself at `top: rect.top - 40` in viewport coordinates and knew nothing
    about the shell. Text under the fixed top bar cannot be dragged — the bar
    takes the pointer — so the first line a member CAN select is the line
    just below it, which is exactly the line whose chip flies up into the
    bar: y=52.1 against a 96px bar, 3,228px² of overlap with
    `header[data-topbar]`, and `elementFromPoint` underneath reporting the
    cluster that carries **Ask Lilith**. An offer to ask Lilith, rendered
    inside the chrome that already offers to ask Lilith. `ShellLayout` now
    publishes `--shell-top-chrome` — the third of the set, after
    `--shell-bottom-chrome` and `--shell-assistant-inset`, and the one
    EVE-VIS-183's row said was missing — from the SAME two constants as its
    own content padding, so the reservation and the padding cannot drift.
    
    **Severity was measured, not inherited.** The obvious move was to call
    this S1 by analogy with EVE-VIS-196. The calibration cell trial-clicks
    every control the pinned-back chip overlaps: **2 covered, 0 refused a
    click.** Confusing, not blocking — S2. The check is the whole difference
    between a severity and a guess.
    
    **It also stopped following the words it was about.** `selectionchange`
    does not fire on scroll and the chip is `position: fixed`, so its
    coordinates outlived the sentence: 349px of drift after one 411px
    scroll, and no way to withdraw when the selection left the screen
    entirely. Now rAF-coalesced re-placement on scroll and resize, and the
    cell asserts the chip moved by EXACTLY what the page moved.
    
    **A green branch that had never executed.** The withdrawal was guarded by
    "if the selection actually left the viewport", and on `/nisaba/scholar`
    it cannot: that document scrolls **411px against a 900px viewport**
    (`probe-8-3-withdraw-and-touch.spec.ts`), so a 4,000px wheel moves the
    chip 114px and stops. The first run's own evidence said
    `selectionOffScreen: false` and the cell passed anyway. Moved to
    `/library` (5,573px of travel), where the selection reaches `top: -3665`
    and the assertion is unconditional.
    
    **Four harness faults that each read as a product defect.** The picker
    measured an element's BLOCK rect and pressed 3px inside its left edge,
    which assumes the words start where the box does; on `/settings` they do
    not, and the press landed on a different `<section>`. It also (a) chose
    `div.srOnly` text — "Keyboard shortcuts: Alt+1 Home…", 60 real characters
    laid out 72×20 at (-1, -242), passing every visibility test CSS offers
    and reachable by nobody; (b) chose collapsible-section subtitles inside
    `<button>`s, which Chromium will not drag-select — eight consecutive
    attempts landed on the words and selected 0 characters; and (c) required
    a 40-character text node, which excluded the whole Nyx room, whose
    longest visible run is **38** ("Royal Observatory, Greenwich · 51.5° N")
    and which selects and offers perfectly well at 38. Each was diagnosed by
    probe before anything was filed: `probe-8-3-settings-selectstart.spec.ts`
    armed `selectstart`/`mousedown` listeners and found `defaultPrevented:
    false` everywhere, which is what ruled the PAGE out. The picker now
    targets the text node's first line box, skips hidden and interactive
    text, and verifies with `elementFromPoint` that the press lands before
    trusting the result; every attempt is recorded with its outcome, so "0
    chars" and "did not land" can never again be confused.
    
    **Calibrated, not merely green** — the standard 8.2 set. One cell pins
    the chip back to the old rule with `addStyleTag` and REQUIRES the same
    measurement to come back red naming the top bar. The five unit locks were
    calibrated by mutating the built component: zeroing `shellChrome()` reds
    only the flip test; removing the scroll listener reds the follow AND the
    withdrawal; removing the viewport-exit branch reds only the withdrawal;
    dropping the margin from the left clamp reds only the width test. No
    guard masks another.
    
    **Colour measured rather than eyeballed** — and it corrected a wrong
    reading. The dusk screenshot looked like a cream chip on a dark page;
    `probe-8-3-chip-contrast.spec.ts` says otherwise: `--l-surface-raised`
    and `--l-ink` both resolve under dusk (`#3a2f1f` / `#f1e8d0`), the
    hardcoded literals are fallbacks that never fire, and the label runs
    **16.13:1** in cream and **10.7:1** in dusk, far above AA. Recorded and
    NOT filed: the hardcoded border is 1.78:1 / 1.37:1 against the chip's own
    fill and the fill is 1.11:1 / 1.29:1 against the page, so what separates
    the chip from the page is its drop shadow and its label rather than its
    edge.
    
    **Three things deliberately NOT claimed.** Touch selection is
    **unverifiable on this harness, not passing**: double tap and a genuinely
    held 900ms long-press dispatched over CDP, under full `isMobile: true`
    emulation, both produced 0 characters — native touch selection belongs to
    a controller headless Chromium does not run — so the cell asserts only
    that no selection yields no offer, and the keyboard select-all cell is
    what carries the narrow-viewport weight. Flipping below the line covers
    the NEXT line of prose, which is inherent to a floating chip with no room
    above and is what every platform's selection toolbar does. And
    EVE-VIS-198 — `/settings` hydrating its billing period one day out
    because the seed invoice is a `Date.now()` frozen at module scope on a
    server that outlived its UTC day — is a real defect this task FOUND but
    does not own; it is ledgered OPEN against the billing seam shared with
    mobile, and the four `/settings` cells allow that one mismatch **by
    name** (`data-profile-invoice-id`/`BillingHistoryList`) so a hydration
    failure anywhere else on the page still reds them.)_
    
  • 8.4 Invocation points: enumerate the registry (assistant-entry.ts, 35+ points) and CLICK EVERY ONE — correct prompt/context arrives in the panel, no dead points, no double-open; visual state of each launcher in situ. _(CLOSED 2026-08-12 s28 — five defects found, four fixed, one ledgered OPEN by name. phase-8-4-invocation-points.spec.ts, 25 cells, over an enumeration that lives in e2e-inspect/support/invocation-sites.ts.

    **A point is a category; a site is a thing you can press.** The registry
    declares 22 points, and `customer-web.inline-help` alone is wired in 14
    files at 20 call sites — most of them on surfaces that exist only in a
    state (a completed Tara session, an active audit run, the midday daypart,
    a previewed search result). So the unit here is the SITE: **35 of them**,
    which is exactly what the task's "35+" was pointing at. The table is held
    to the source by a totality cell that mines every
    `openAssistantInvocation('…')` in `apps/oshun/web/src` and requires the
    per-file counts to match what the table claims to cover — calibrated by
    dropping one file's count from 3 to 2, which reds it naming the file.
    
    **Three layers per press, because "the context arrived" can be true at
    one and false at the next**: the LAUNCH (exactly one
    `oshun:assistant-open`, carrying the point's entry source and the seeded
    prompt), what the SERVER is told (the `contextHandoff` on the session and
    turn POSTs), and what the MEMBER is shown (the panel's own "You opened me
    from …"). One press → one launch held in **all 27** clicked and keyed
    sites. The single multi-fire is `/assistant`'s arrival EFFECT — React's
    development StrictMode double-invoke, the same mechanism session 25 pinned
    at 2/2/2 against 1/1/1 (EVE-VIS-190) — so that cell asserts what a double
    cannot excuse (one source, one prompt, one panel) and records the count
    rather than filing it.
    
    **EVE-VIS-199 (S1) — the launcher that could not be read, and the room it
    stands in.** The Tara completion follow-up measured **15.42:1 in cream and
    1.09:1 in dusk with the same computed colours**, which is only possible if
    the surface moves the other way from the theme. It does: the immersive
    plate is `rgba(var(--l-ink-rgb), 0.97)`, and under dusk the cinema turned
    into a white sheet under white text. Walking every text node then found
    the same class in CREAM, where nobody had looked — the global `h1–h3` rule
    colours a heading `--l-ink`, the plate's own BACKGROUND token, so the
    session title sat at **1.07:1 in both themes**. Player **19 of 22 nodes
    below AA in cream, 13 of 22 in dusk**; completion 14 below AA in cream.
    After: **0 and 0, in both themes**, worst node 5.89:1.
    
    **EVE-VIS-200 (S2) — the card that promised the ask and sent nothing.**
    `/search`'s preview card is headed "Assistant context — preserve the
    current result while you refine the next move", prints the question in
    full, and names its button from the result ("Compare with Lilith"). All
    four controls launched with **no prompt at all**. The intent was not
    missing, it was written down and never sent: `openPreviewAssistant`'s
    dependency array names `activePreviewItem` and its body ignores it.
    
    **EVE-VIS-201 (S3) — the registry's own truth, in three places.**
    `customer-web.header-button` was wired to nothing while carrying member
    copy, a launch curation and an anchor tie; the launcher that really serves
    shell-less surfaces opened the assistant OUTSIDE the registry, with an
    entry source no copy could describe (so the panel's opening sentence lost
    its clause on every such page); `customer-web.empty-state-cta` was
    likewise declared and unwired. And the graph's two anchor ties were
    exactly backwards — read off the DOM, the shell's button carries
    `shell.assistant-trigger` and the floating one carries
    `assistant.launcher`. **Dead points: 2 → 0.**
    
    **EVE-VIS-202 (S2) — the panel told the member the truth and the server
    got `inline-help` for everything.** The bus listener re-pointed every
    launch at the inline point and corrected only the member-visible half with
    `entrySourceOverride`, so `contextHandoff.launchIntent.invocationPointId`
    disagreed with the `entrySource` beside it — in the field an audit log
    attributes by. `assertLaunch` now requires every handoff source seen on a
    press to equal the point pressed, in all 14 press cells.
    
    **EVE-VIS-203 (S2), ledgered OPEN and named.** On the surfaces the shell
    does not own — the three Arete workspaces, the Nyx event page, the reader
    — `AssistantHost` passes no `contextHandoff` at all: the prompt lands, the
    room does not ("You're in Lilith." against "You're in Nyx (Night Sky)."
    one level up), and the BFF never learns which room the launch came from.
    Not fixed inside 8.4 on purpose: the fix is a shared handoff builder both
    hosts call, and a half-built one — right domain, wrong memory scope —
    would be worse than none, because `memoryScope` is a privacy fact the
    member is shown. Checked before claiming more: `permittedToolGrants` are
    shape-validated by the BFF and enforced nowhere, so that half is latent.
    
    **Four things deliberately NOT claimed.** Veritas's story explainer is
    real code on a room this release cuts — `/domains/veritas/story` redirects
    to `/release-scope/veritas`, asserted rather than clicked. Studio's Copilot
    below its 768px minimum renders `disabled` with
    `data-assistant-invocation-state="blocked"`, which is the guard answering
    correctly, not a broken control. `/assistant`'s in-page button at 390 is
    covered by the panel the route opens on arrival — the launcher is under
    the thing it would have opened. And `admin-web.*` (5 points) and
    `customer-mobile.*` (4) belong to Phases 9 and 10; the cell records them
    as owned elsewhere rather than skipping them silently.
    
    **Five harness faults, each caught before it became a defect report.** A
    shared browser context carried `/assistant`'s open dock into the next
    route and reported three launchers "that render but cannot be pressed" —
    the sweep now takes a fresh context per route, and a control the app marks
    blocked or that the assistant panel itself covers is recorded as such
    rather than counted as unreachable. A flat 2 s wait after "Start the
    audit" read a launch that fires on the run POST's 201 as a dead point, and
    a single `count()` read the board's loading phase as a launcher that
    renders nothing (it waits for the control, or for the documented
    alternative that replaces it, before calling anything absent). The
    palette's first row for "ask assistant" is `nav-assistant`, which ROUTES
    to `/assistant` — pressing it produced a `customer-web.deep-link` launch
    from the entry route's own effect, not a palette launch; the cells address
    `assistant:ask` by id and record that the palette lists the command and
    its ⌘J shortcut as two identically-worded rows. **And a landmine worth its
    own line: killing `next dev` mid-life and restarting it left a `.next`
    cache serving `/api/auth/session` as 404**, so every seeded member was
    bounced to `/welcome?redirect=…` and `/search` looked like it had lost its
    controls — server-side the same cookie fetched the page at 200 with the
    controls in the HTML. `rm -rf .next` and a cold start; the discriminator
    is that the failure is CLIENT-side only.)_
    

Phase 9 — Builder tier: admin Copilot, confirm cards, workbench UI#

  • 9.1 Admin drawer (post-1.6) full five-lens pass: header, workspace/ operator/memory chips, transcript, input, meta inspector collapsed state — desktop + narrow, light + dark (admin theme story?). _(CLOSED 2026-08-12 s28 — one S2 found and fixed, EVE-VIS-204. phase-9-1-admin-drawer.spec.ts, 4 cells.

    **1.6 cleared the room and the conversation never moved into it.** The
    drawer is 900px tall and the copilot got **107px** of it. `.panel` was a
    grid with `auto auto auto auto auto 1fr` over FOUR children, so the `1fr`
    landed on a phantom row after the last one: rows measured `79 / 64 / 107 /
    28 / 0 / **622**` — the tallest region in the drawer held nothing at all —
    while the transcript carried a flat `maxHeight: 260`. Driven with four
    exchanges, the operator scrolled a 260px window (`scrollHeight 496`) with
    372px empty below it. Now a flex column whose conversation claims the
    spare height: **728px of 900**, transcript 632px, nothing overflowing.
    
    **The five lenses, measured on each named part.** Layout: every part's
    rect inside the panel's, the panel's inside the viewport. Overlap:
    `elementFromPoint` at each part's centre plus a trial click on the close
    button, the input and the disclosure summary. Scroll: which container
    scrolls (the transcript, and only it — the panel itself does not
    overflow). Contrast: every visible text node in the drawer against its
    composited backdrop — **11 nodes, worst 6.96:1** once the walker stopped
    counting text it could not see. Copy: each part's strings recorded.
    
    **The theme question the task asked, answered rather than assumed.** The
    admin app declares `color-scheme: dark` and ships ONE palette; there is no
    switch and no `prefers-color-scheme` rule in its sources. So "light +
    dark" here can only mean theme-INVARIANCE, and that is a measurement: the
    drawer is sampled under both `emulateMedia({ colorScheme })` values and
    every colour must match. It does.
    
    **Narrow is a refusal, not a drawer.** `admin-web.assistant-panel`
    declares `minViewportPx: 1024`, so at 390 the header's Copilot renders
    `disabled` with `data-assistant-invocation-state="blocked"` and no panel
    mounts. The cell asserts the refusal the app states in the DOM rather than
    pretending to inspect a drawer that cannot open. Recorded, not filed: the
    disabled button carries no `title`, so the operator is told nothing about
    why — the guard knows (`viewport-too-small`, `minViewportPx: 1024`) and
    the button does not say.
    
    **One instrument fault, fixed before it became a finding.** The contrast
    walk first reported four disabled persona chips at 3.93:1 — inside the
    COLLAPSED inspector. Chrome hides a closed `<details>` with
    `content-visibility`, which leaves its children's boxes measurable, so the
    walker was reading text nobody can see; it now asks
    `Element.checkVisibility()`. The one remaining sub-AA node is the Send
    button at 2.71:1 **while disabled** — Chrome's own `graytext`, which WCAG
    1.4.3 exempts and the app never styles. Measured enabled: **17.09:1**, and
    that is the state the cell asserts.)_
    
  • 9.2 Confirm cards on admin: every workbench write tool triggers its card with exact-action copy — create*work_item, update_work_item (+transition), open_thread, comment_thread, link_graph_refs, draft_decision, transition_decision, create_feature_proposal, create_content_brief, dispatch_content_brief. For each: Approve lands (row verified), Decline is a true no-op, outcome message copy, card removal, TTL expiry rendering. *(CLOSED 2026-08-12 s29 — four S2/S3 found and fixed across the two halves (EVE-VIS-205 s28; EVE-VIS-206/207/208 s29), plus EVE-VIS-209 ledgered open for 9.3.

    **The drawer half (s29), live at the real admin copilot.**
    `apps/oshun/web/e2e-inspect/phase-9-2-admin-confirm-cards.spec.ts`, 3
    cells, each driving a real model turn that parks a real action and
    reading the answer out of Postgres rather than out of the drawer: the
    card's five lenses + APPROVE lands the row (`priority high`, `status
    draft`, read back by title); DECLINE leaves no row and says so naming the
    row; and an expired card, driven by shortening the deadline ON THE WIRE
    so a real action really expires in the browser.
    
    **EVE-VIS-206 (S2): the card offered Approve forever and answered the tap
    with our own words.** The deadline has travelled with the card since
    EVE-VIS-100 — the admin drawer read three fields off the frame and
    dropped it, so it had no way to know the card had died. The operator
    pressed a dead control and read `The action could not be executed (Action
    not found, expired, or already resolved). Nothing was changed unless the
    assistant confirms otherwise.` — three possibilities, none chosen, and a
    hedge about the only question they have. The sentences now live in
    `@oshun/shell-assistant`, so "did my data change?" cannot be answered two
    ways by two surfaces; the member's own copy spec still passes unchanged
    against the lifted implementation. Looking at the screenshot then caught
    the fix's own residue: an expired card still promising "nothing happens
    until you decide".
    
    **EVE-VIS-207 (S2): four cards named half the action.** `update_work_item`
    with fields AND a move printed only the move; `link_graph_refs` printed
    the survivors of a REPLACEMENT and never what it dropped;
    `dispatch_content_brief` never mentioned the draft→ready walk it makes;
    `transition_decision` retired an ADR without naming its successor. The
    case table is now TOTAL over the call — every non-identifier argument is
    either named on the card or exempted with a reason, so an argument nobody
    thought about fails the tool's cell instead of vanishing.
    
    **EVE-VIS-208 (S2): one approved card, two transactions.** A draft item
    cannot go straight to `ready`, so "rename this and mark it ready"
    committed the rename, hit the lifecycle, and reported "nothing was
    changed" over a renamed row. `editWorkItem` now does both in one
    transaction with legality checked first, and the card refuses an
    impossible move while the turn is still running.
    
    **Recorded, not filed:** the model says "confirm it on the action card"
    while the operator's button says **Approve** — one control, two words,
    which is 13.2's terminology table rather than a defect. And the model
    does not always call the write tool when asked to capture something — the
    recon needed a second phrasing twice — so each cell records `asks` beside
    its measurements (1 on the closing run) instead of retrying it away
    invisibly.
    
    **The server half (s28).**
    `apps/oshun/bff/src/workbench/confirm-cards.integration.spec.ts`, 24 cells
    at the real dev Postgres, drives EVERY mutating binding through the real
    toolset and confirm bridge — the case table is checked against the
    bindings themselves, so a write tool added without a case reds the file
    rather than shipping unconfirmed. Per tool: the card's copy, nothing
    written before the decision, APPROVE lands the row (read back out of
    Postgres), a second approve refused, DECLINE writes nothing and cannot be
    undone, and an expired card cannot be approved.
    
    **EVE-VIS-205 (S2): five cards named a machine id and nothing else.** The
    member side settled this in EVE-VIS-098 — "the card is the authority, so
    the card is what has to name the row" — and made `confirmSummary` async so
    the title could be read back by id. The ADMIN binding type was left
    synchronous, so the workbench's cards could only echo the model's
    arguments: `Update work item wi-64826f77-4434-406…` named neither the item
    nor the change, and `Post a comment to thread th-…` named neither the
    thread nor the comment. An operator cannot judge those, and past that card
    the write happens. Now: `Update “…” — title`, `Comment on “…”: “…”`, `Link
    “…” to tara`, `Move ADR “…” from draft to proposed`, `Dispatch “…” to
    docs-center as …`. Reading the row at card time also means an id the model
    INVENTED refuses while the turn is still running rather than after the
    tap.
    
    Its case table has since grown to 32 cells (the four combined-effect
    cases and EVE-VIS-208's two atomicity cells).
    
    **The five lenses on the card, measured.** Layout: the card's rect inside
    the drawer's and the drawer's inside the viewport. Overlap:
    `elementFromPoint` at the card's centre and at each button's, plus a
    trial click on both. Scroll: which container in the drawer actually
    overflows (none, with a card parked) and whether the whole card is inside
    the conversation region — a gate you cannot see is a gate you cannot use.
    Contrast: every text node in the card against its backdrop, worst
    **14.58:1**. Copy: names the row, no id-shaped token, states the tool and
    that nothing has happened yet. Console clean throughout.
    
    **EVE-VIS-209 (S3, open) found in passing:** the admin transcript prints
    the model's markdown literally — `**Task: "…"**` — where the member panel
    renders it. Left for 9.3, which inspects this same transcript.)_
    
  • 9.3 Read tools in conversation: list*work_items (with nodeId closure filter question phrasing), get_work_item history rendering, list_decisions, read_thread WITH its grounding block — verify the model cites ids and the grounding is used, not parroted. *(CLOSED 2026-08-12 s29 — four fixed: EVE-VIS-209 (S3), 210 (S2), 211 (S3), 212 (S3). phase-9-3-read-tools.spec.ts, 2 cells at the real drawer.

    **The corpus is the instrument.** Over an empty intent plane every read
    tool succeeds and shows nothing, which reads like a working happy path
    and is not one. `apps/oshun/bff/scripts/seed-workbench-corpus.ts` writes
    one whose SHAPE makes each answer checkable: work on `tara` AND on `nyx`
    (so a tara answer has something to EXCLUDE), one item linked only to a
    tara DESCENDANT (so the containment closure is doing the finding, not a
    string match), one rejected item (so "open" has something to leave out),
    one item carried through four events (so the history has depth), two ADRs
    and a thread anchored to a real node. It writes through the real store,
    so every row is a ledger event; it is idempotent and tagged for 9.6.
    
    **EVE-VIS-210 (S2): a discussion had no way in.** `read_thread` and
    `comment_thread` take a thread id and nothing produced one — a thread was
    reachable only inside the conversation that opened it. Asked to summarise
    a discussion by subject, the copilot said "I don't have any discussion or
    graph nodes to draw from on that topic" while the thread sat in Postgres.
    `list_threads` now exists (anchor-closure filter, most recently active
    first, honest cap), and the lock is a DISCOVERY WALK — list→open for all
    three entity kinds using only ids the lists handed back — so a future
    entity with a reader and no lister fails there.
    
    **EVE-VIS-211 (S3): `wi-d601152c` is not an id.** Three of five recon
    answers cited the first uuid segment; the prompt said "cite item ids" and
    a shortened one still looks like obedience. Held as a property, not a
    sentence: every `wi-`/`adr-`/`th-` token in every reply must resolve to a
    row, which a truncated id fails exactly as an invented one does.
    
    **EVE-VIS-212 (S3): the operator was answered like a member** — "I can
    walk you to Nisaba, or we can pick up a Tara session you had going", said
    in the admin console. The builder prompt now states who the reader is.
    
    **EVE-VIS-209 (S3, opened in 9.2): closed here.** The member panel's
    closed-subset markdown renderer moved to `@oshun/shell-assistant/markdown`
    — same parser, its 44-cell spec green against the new path — and the
    admin transcript renders assistant messages through it. Live: 22
    `<strong>`, 11 `<li>`, zero literal `**`.
    
    **What was measured on the answers themselves:** every tara-closure item
    named (including the descendant-linked one), the nyx item and the
    rejected one absent, the ledger's four events reflected with the human
    who made them, the ADR named with its status, the grounding USED (cites
    `tara.courses`, `tara.courses.progress`, `/activity`) and not PARROTED
    (none of the block's own header or edge labels), every cited graph node
    real, replies inside the transcript with no horizontal overflow, the
    newest answer in view, worst contrast above AA, console clean.)_
    
  • 9.4 /assistant/graph explorer — full five-lens inspection of all four lenses at real data scale (the SVG tree with 747+ flows: scroll performance, row legibility, coverage colors in both themes), inspector on 10 diverse nodes (domain/feature/flow/step/ route/journey), overlay toggles, filters (unverified/waived/ open-work), deep links from open_graph_explorer land with the right lens+filter+focus, builder-scope gate + retry, error state. _(CLOSED 2026-08-13 s29 — three fixed (EVE-VIS-214 S1, 215 S2, 217 S3) and one filed OPEN for the user (EVE-VIS-216 S1, a product decision). phase-9-4-graph-explorer.spec.ts, 5 cells, every assertion comparing the board against the BFF's own three payloads read beside the browser.

    **EVE-VIS-214 (S1): the board asked before it had a credential.** The
    three fetches read the auth context's token mirror on mount, and that
    mirror is empty until `/api/auth/session` answers — so on three arms
    (anonymous, member, member with an `admin:*` cookie) all three requests
    went out bare and came back **401 `missing_bearer_token`**, and every one
    of them was shown "Builder scope required. Sign in with an operator
    account." A sentence about the wrong door. It is the audit board's
    EVE-VIS-184 on the sibling surface; same fix, same shape.
    
    **EVE-VIS-216 (S1, OPEN — for the user):** with that race fixed, the
    honest answer is a permanent refusal. `buildAuthSession` hard-codes every
    member access token to `scopes: ['domain:*']`, and two of the explorer's
    three endpoints require `admin:*` — so no member-web sign-in can open the
    graph explorer, and the deep link the admin copilot hands an operator
    lands on a page they cannot pass. The DATA is fine (86 rows, 4 lenses,
    correct coverage with an operator bearer attached); the identity is
    missing. Three ways out, all product decisions: an operator sign-in on
    the member web, or move the explorer to the admin console where
    `oshun-admin-session` already exists, or serve the two builder endpoints
    through an admin same-origin proxy. Not chosen here.
    
    **EVE-VIS-217 (S3): the lens spoke only in colour.** Verified / unverified
    / waived was a stroke colour and nothing else (WCAG 1.4.1), and the
    inspector — which was handed `traceability` and destructured it away —
    never said which. Rows now carry `✓ / ✕ / ~` beside the kind glyph and
    the inspector states the verdict in a sentence, both from the analyzer's
    own lists.
    
    **What the five lenses found, measured:** the tree draws exactly the
    curated domains+features+flows the catalog holds (**86 rows**; steps live
    in the inspector by design); every flow's colour AND character agrees
    with the analyzer's `unverifiedFlowIds` / `waivedFlowIds`, never
    recomputed; row labels at 12px/11px with worst contrast **5.74 (cream) /
    6.50 (dusk)**, all above AA; the tree is 2,244px in a 900px viewport and
    reachable by page scroll with the SVG's own container the only sideways
    scroller; the inspector on ten nodes names each node, lists only ITS
    linked work, and stays inside the viewport; the three filters reproduce
    the analyzer's sets exactly (set equality, not counts); turning the
    overlays off removes every badge and every coverage colour; the work
    kanban's column counts match the board's rows; the ADR timeline and the
    manifest table name every row the server sent; and `?lens=…&filter=…&
    domain=…` lands on all three at once with no foreign domain in the tree.
    Console clean.
    
    **Two premises in the task's own wording, corrected against the data:**
    the explorer's `feature-catalog` serves the CURATED catalog — 19 domains,
    35 flows, 725 journeys mapped — not "747+ flows", which is the whole
    product graph; and route/journey nodes are not in that payload at all, so
    "10 diverse nodes" is domain/feature/flow (steps render inside the
    inspector, which is where the task's own component puts them).)_
    
  • 9.5 Workbench board data views: work kanban columns at real scale, ADR timeline with a real supersedes chain, changes lens manifest diff with a pasted previous manifest (happy + malformed-paste error copy). _(CLOSED 2026-08-13 s29 — two S2 found and fixed (EVE-VIS-218, 219), and two INSTRUMENT faults in this harness's own contrast walker found and fixed with them. phase-9-5-board-data-views.spec.ts, 3 cells.

    **EVE-VIS-218 (S2): the changes lens diffed against anything that
    parsed.** Paste `123`, or `[1,2,3]`, or `{"hello":"world"}`, and all
    eleven sections of the artifact were reported as **"new section"** — a
    promotion history stated as fact about an input that contained none; a
    `sections` array without counts printed **"changed (NaN nodes, NaN
    edges)"**. Only JSON syntax was guarded. Now the SHAPE is checked and the
    paste is refused by name ("that JSON is an array, not a manifest object",
    "it has no \"sections\" array", "\"curated\" is missing sectionHash,
    nodeCount, edgeCount"), with no diff column until a real manifest
    arrives.
    
    **EVE-VIS-219 (S2): the status palette failed AA as text — and the
    instrument could not see it.** The tree's work badges (`⚑2`, amber on
    cream) measured **2.62:1**, the ADR timeline's `superseded` the same, and
    `accepted` **4.15:1**. Two harness faults hid the first: the contrast
    walker read `style.color` on SVG text, which is painted by `fill`, and it
    skipped text nodes of length ≤1 — which is exactly what JSX makes of
    `{'⚑'}{count}`. So 9.4's clean pass over 107 nodes was a statement about
    a tree it had half entered. Both fixed; the same walk now measures **246
    nodes** and the badges read **5.74:1**. The colours themselves moved off
    three hard-coded hexes onto the design system's `--l-ok` / `--l-alert` /
    `--l-warn`, which the stylesheet already tunes per theme — and 9.4's spec
    now resolves those tokens from the page instead of copying hexes.
    
    **What was measured.** Kanban: every status column carries its real count
    (`draft · 4`, `triaged · 1`, `ready · 2`, `verified · 1`, `rejected · 1`),
    every live item is on the board by title, an unlinked item says
    "unlinked" rather than showing an empty swimlane, and all text clears AA.
    At scale — 240 rows put in at the HTTP boundary, synthetic and labelled
    as such, because the question is what the RENDERER does — 10 columns
    survive, 240 cards render, the tallest column is 4,040px and reachable by
    page scroll, the board scrolls sideways to its later columns, and **zero
    cards are clipped**. (The first version of that check found 1 column and
    0 cards and "passed" — a check over an empty set is not a check; it now
    asserts the card count before asserting anything about clipping.)
    ADR timeline: the seeded chain — two ADRs walked proposed → accepted, the
    older then retired — renders with `supersedes → adr-…` on the SUCCESSOR,
    which is where the store records it, the retired one showing
    `superseded`, and every decision the board holds present.
    Changes lens: a previous manifest built from the one the server just sent
    yields the exact deltas (`changed (+7 nodes, -3 edges)`, `new section`,
    `unchanged`), and clearing the box returns the table to its plain
    state.)_
    
  • 9.6 Cross-check the intent plane after all Phase-9 writes: ledger rows all actor-attributed, no orphan test debris left behind (clean up created items honestly — reject/park with notes, never DELETE rows). _(CLOSED 2026-08-13 s29 — one S2 found and fixed (EVE-VIS-220). apps/oshun/bff/scripts/crosscheck-intent-plane.ts asks the three questions against the rows and can act on the third.

    **Attribution: clean.** 96 ledger events at the time of the check, nine
    distinct actors — `human:operator-studio-01`, `assistant:assistant:
    user-elikem`, `coding-agent:claude-code-local`, two artifact-verifier
    ids, and the rest — and **zero anonymous events**. The store refuses an
    anonymous write; this proves nothing went around it.
    
    **EVE-VIS-220 (S2): `replay(ledger) == rows` was FALSE, and nothing said
    so.** The plane's central invariant is proved by a spec that wipes the
    tables first, so it had never been asked of the live plane. Asked here:
    13 work items and 18 decisions/threads existed in the replay and not in
    the tables — the residue of fixture cleanups that delete the projection
    row and keep its events, which is exactly what this task's own rule
    forbids. Every such cleanup now takes the history with it (`forgetEntity`
    in three integration specs), and the accumulated residue is repaired once
    by `--repair-orphans`, which removes events for entities the projection
    no longer has and nothing else. After: `{workItems: true, decisions:
    true, threads: true}`, zero replay-only ids.
    
    **Debris: retired, not deleted.** Five harness rows were found (the 9.2
    drawer's approved row, the `polish-9-3` corpus). Four were still sitting
    in the backlog as draft/triaged/ready; all four are now **parked with a
    note** naming the task that created them and saying they are evidence
    rather than work — through the store, so each parking is itself an
    actor-attributed ledger event. Nothing was deleted. Re-running 9.3 and
    9.5 against the parked corpus keeps both green, which is the check that
    the retirement did not quietly change what those tasks measure.
    
    **Recorded, not filed:** `intent-store.integration.spec.ts` wipes the
    WHOLE intent plane (rows and events) at boot — deliberately, since dev
    projection tables are disposable — so `pnpm test:integration` in the BFF
    erases the corpus. `seed-workbench-corpus.ts` is idempotent and rebuilds
    it; that is the recovery, and it is worth knowing before wondering where
    the fixtures went.)_
    

Phase 10 — Mobile & PWA assistant surfaces#

  • 10.1 Emulator pass (Android per the toolchain recipe): assistant sheet open/close, buffered JSON turns render, structured navigateTo pushes the right screen, transcript scroll on touch, keyboard avoidance. _(2026-08-13 s30: CLOSED. Five legs driven on a real arm64 API-35 emulator (Pixel 6, 1080×2400, headless) against the live BFF — and two of the five were completely broken, both S1, both now fixed and verified on the device.

    **The infra had to be rebuilt from nothing first** — no AVD existed and
    `system-images/android-35/google_apis` had no ABI directory. Recipe
    additions for the next session: Google's CDN throttles WITHIN a
    connection (a fresh conn pulls 584 KB/s and decays to 23 KB/s), so the
    July chunked-append downloader now projects ~21 hours; a PARALLEL
    downloader with one part file per chunk (never a shared file — that was
    July's corruption mode) finishes the 1.78 GB image in ~40 min. Two
    Gradle builds then failed for environmental reasons worth knowing:
    `dl.google.com: nodename nor servname provided` because the download's
    workers had starved the resolver, and a silent wedge on stalled HTTP
    connections that only cleared with
    `-Dorg.gradle.internal.http.socketTimeout=60000`. Third build:
    SUCCESSFUL in 12m 8s, 93 MB APK.
    
    **Leg 2 — "buffered JSON turns render" — was failing on EVERY build, and
    is the headline: EVE-VIS-221 (S1, FIXED).** `POST /v1/assistant/sessions`
    answered **401**, so session creation threw, `deliverLiveReply`'s outer
    catch fired, and every member turn silently became the on-device offline
    model. The live BFF path S11 exists to guarantee had never run for a
    real member on mobile. The contrast is what makes it airtight: from the
    SAME device in the SAME session, `/auth/signup`, `/auth/sessions`,
    `/auth/refresh`, `/v1/billing/subscription-management`,
    `/v1/feature-flags/evaluate` and `/v1/notifications/preferences` all
    returned 200 — only the assistant 401'd, because all four production
    `createOshunBffClient` call sites authenticate from
    `EXPO_PUBLIC_OSHUN_DEV_AUTH_TOKEN`, which `runtime.ts:39` defaults to
    `''`. **Why no gate caught it:** the one e2e covering mobile assistant
    turns waits on
    `mobile-assistant-turn-history-entry-assistant-(live|offline)-prompt-.*`,
    which the OFFLINE branch satisfies — it passes whether or not the agent
    is ever reached. Fixed by asking the member's session first (dev env var
    demoted to a harness fallback), in the sheet, the offline replay queue
    (whose handlers also skipped on `!devAuthToken` and reported the skip as
    success), the safety journey and idea capture. **Verified with the dev
    token REMOVED from the bundle**: session `201`, `/turns` `200`, no
    offline notice, and a real structured `navigateTo` push into Nyx.
    
    **Leg 5 — keyboard avoidance — was 100% broken: EVE-VIS-222 (S1,
    FIXED).** Tapping the prompt field hid the prompt field. Measured IME
    frame `[0,1517][1080,2400]` while the app window stayed
    `[0,0][1080,2400]` — **unresized**, because `edgeToEdgeEnabled=true`
    makes Android deliver insets rather than resize, so the manifest's
    `adjustResize` lifts nothing and no `KeyboardAvoidingView` exists
    anywhere in the app. The composer's bounds were byte-identical before
    and after the IME opened (Send at y=2197..2313, i.e. 680px below the
    keyboard's top edge); typing still landed in the invisible field and
    still enabled the invisible Send. After the fix the composer sits at
    y=1374..1494, clearing the 1517 edge. **Maestro cannot see this class of
    defect and the suite already worked around it** — a view under the IME
    stays in the hierarchy with on-screen bounds and `clickable=true`, so
    `assertVisible` PASSES; `assistant-voice-text-persona-switching.yaml`
    calls `hideKeyboard` with a comment saying the keyboard covers the send
    button. Geometry is the only honest witness, which is why the lock
    compares against the `type=ime` inset frame.
    
    **Legs 1 and 3 are clean.** The sheet opens from the launcher and closes
    on the backdrop; and "take me to nyx" produced an object `navigateTo`
    that closed the sheet and pushed the real Nyx room ("Resolved screen:
    Nightly highlights"). Worth recording for later phases: 5/5 sampled
    prompts returned a STRUCTURED `navigateTo` whose domain is in
    `MOBILE_NAVIGABLE_DOMAINS`, so this path is reliable, not lucky.
    
    **Leg 4 scrolls on touch — and that is the only good thing about it.
    EVE-VIS-223 (S1) is OPEN and gates the phase.** At open the sheet shows
    no conversation at all, not even the assistant's own greeting;
    `mobile-assistant-turn-history` is absent from the accessibility tree
    entirely. It takes **five 800px swipes** (~4 screenfuls of the sheet's
    1182px scroll body) to reach the transcript's top edge, under six blocks
    of diagnostic chrome. Sending changes nothing on screen: there is no
    `scrollToEnd` anywhere in the file, so the reply lands below the fold and
    the only evidence anything happened is a counter inside the SAFE
    FALLBACK panel moving `1 transcript turn` → `3`. Half of this is
    mechanical (auto-scroll, which web has along with "Jump to latest"); the
    other half — that the conversation sits beneath Transcript state, Mode
    state, Avatar mode, Persona handoff and Safe fallback — is an
    information-architecture decision on a surface whose panels have testIDs
    and e2e coverage, so it is flagged rather than redesigned unilaterally.
    
    **Also ledgered from this pass:** EVE-VIS-224 (S2) the buffered turn
    carries the STREAMING client's 30s number as a TOTAL budget — web resets
    its timer per chunk, mobile arms one abort before `fetch`, so a 63s turn
    is fine on web and dead on mobile; measured 5 of 8 turns over the limit,
    median 38.4s, all 200 server-side, and reproduced live on the device.
    EVE-VIS-225 (S2) nothing indicates a reply is coming during that wait.
    EVE-VIS-226 (S2) `/v1/mobile/diagnostics` 404'd ×9 — the app posts crash
    telemetry to a route the BFF does not serve. EVE-VIS-227 (S3) a doubled
    slash in `Path //nightly-highlights` and a room label breaking mid-word
    as `CURR`/`ENT`/`ROOM`.)_
    
  • 10.2 Mobile voice: mic/speaker sheet UI (standing "pending emulator pass" item) — record → transcript → turn; TTS playback states. _(2026-08-13 s30: CLOSED, and the honest answer is that the surface does not exist — so what this task could actually test is whether the sheet SAYS so truthfully. It did not, and that is now fixed.

    **There is no voice UI to drive.** No mic control, no recording state, no
    TTS playback control anywhere in the assistant sheet: a hierarchy dump of
    the whole open sheet on the emulator matched **no** testID containing
    mic/record/voice/speak/tts/audio. `transcribeAssistantAudio` exists on the
    mobile BFF client (`oshun-bff-client.ts:754`) with **zero callers** in
    `src/` or `app/`. `microphone: false` and `syntheticVoice: false` are
    hardcoded literals in two places in `MobileAssistantSheet.tsx`. So
    "record → transcript → turn" and "TTS playback states" have nothing to
    exercise. **This is NOT a fabrication stub** — the sheet declares the
    capability unavailable instead of pretending to listen, which is the
    honest fail-loud shape, and it is the right call while the path is
    unwired.
    
    **What IS testable — the refusal — was measured on device.** Tapping the
    sheet's Voice option correctly refuses and stays in Text
    (`Voice-first: Inactive`), rendering
    `Voice is unavailable (missing microphone). Continuing in Text.`
    alongside `Mic: Missing` and `Voice: Missing`. Calm, correctly shaped,
    and not the raw-jargon `not-allowed-in-shell` branch — `voice-first` is
    allowed in the `customer` shell, so this is the clean downgrade path.
    
    **EVE-VIS-229 (S2) FIXED — the sheet disclosed a voice it does not
    have.** The transcript-state chip read
    `Disclosure: AI assistant + Session memory + Synthetic voice` while two
    capability chips in the same sheet read `Missing`. Cause: two hardcoded
    literals inside ONE component disagreeing —
    `syntheticVoiceEnabled: true` at :787 against `syntheticVoice: false` in
    the capabilities. Since a disclosure strip states what the assistant is
    actually doing, this was a false disclosure, not a cosmetic slip. It now
    derives from the same capability the chips read, so the two cannot
    diverge again. Verified on device: the chip now reads
    `Disclosure: AI assistant + Session memory`.
    
    **EVE-VIS-228 (S2) LEDGERED, OPEN — the refusal names the wrong cause.**
    "missing microphone" is false: the phone has one, and the app already
    records with `expo-audio` elsewhere (`meditationPlaybackEngine`,
    `mobileCapturePermissionsBridge`). The true cause is that the
    assistant's voice path was never wired, and `microphone: false` is a
    literal rather than a capability query — so the member is told their
    hardware is missing and will go hunting an Android permission that would
    change nothing. Not fixed here because the honest fix needs a NEW
    downgrade reason in the shared `interaction-modes` vocabulary (something
    like "not available on this surface"), which is a `libs/oshun/
    shell-assistant` change that reaches the web shell too — worth doing
    deliberately rather than as a mobile side effect.)_
    
  • 10.3 PWA at phone viewport in Chrome: panel usability, tour player on touch, SelectionAsk with touch selection. _(2026-08-13 s30: CLOSED. Six cells green — three legs × cream and dusk — at a real 390×844 touch context (hasTouch AND isMobile, deviceScaleFactor: 3), driven by TAP rather than click. Spec: e2e-inspect/phase-10-3-pwa-phone-touch.spec.ts; evidence in e2e-inspect/.evidence/phase-10-3/.

    **Panel usability — clean.** Launcher and composer are reachable by
    thumb, judged with `tap({ trial: true })` rather than `isVisible()`
    (present ≠ pressable — EVE-VIS-026 and EVE-VIS-196 are both "a control
    that renders and cannot be pressed"). No horizontal overflow on `/`, the
    composer stays inside the viewport, and `pageErrors + appErrors` are
    empty. A separate probe scrolled every scroller in the panel to its end
    and found **no content left unreachable under the pinned composer**.
    
    **Tour player on touch — clean.** The tour advances on TAP in both
    themes, and the card fits at every step (`overflowBottom` −116, −24 and
    similar; `scrollHeight === clientHeight`, so it is not internally
    clipped either). **Calibrated**: aiming the advance at a non-advancing
    element turns both cells red with "tour step did not advance on tap", so
    the leg can fail.
    
    **SelectionAsk — split verdict, and the split is honest.** The PRODUCT
    half is clean: given a selection at phone width, the offer appears, is
    tappable by thumb, and sits inside the 390px bounds (EVE-VIS-197 was this
    control drifting out of position). The GESTURE half **cannot be driven in
    this harness at all** — calibrated before asserting anything:
    `probe-8-3-withdraw-and-touch` under `isMobile: true` returns
    `0 chars / 0 chips` for BOTH a double tap and a CDP-held long press, on a
    different page. Chromium's touch-selection controller does not raise a
    selection here, so an assertion on the gesture would have measured
    Chromium and filed a false defect saying SelectionAsk is broken on touch.
    Carried to 10.4 as an honest instrument gap, not a product verdict.
    
    **Two of my own screenshot readings did NOT survive measurement, and
    that is the lesson worth keeping.** The panel capture looked like the
    "ACROSS THE APP" card was clipped under the composer — measured, only one
    element overlaps the composer's rectangle and `elementFromPoint` says the
    composer is not what is painted there, so nothing is occluded. The tour
    capture looked like the narration was cut off — measured, the card
    clears the viewport bottom at every step. At `deviceScaleFactor: 3` a
    screenshot's own edge is indistinguishable from a clip; geometry decides,
    not the eye.
    
    **One real defect confirmed, and it is already open: EVE-VIS-124.**
    `/nisaba/scholar` — the page the selection leg runs on — reports
    `document.scrollWidth` **773** against a 390 viewport, i.e. **383px of
    horizontal overflow** from fixed 320px columns that never collapse. Under
    mobile emulation the page widens its own layout viewport, so
    `window.innerWidth` reads 773 there while `/` reads 390 in the SAME
    context. Measured into that row rather than filed again; it belongs to
    the Phase-12 surfaces sweep by that row's own note. The 10.3-specific
    consequence is recorded there too: the passage a member selects runs
    off-screen, so they cannot read what they are asking Lilith about.
    
    **Harness landmine re-learned (EVE-VIS-104's class):** screenshots must
    NOT be written under `.artifacts/`, which is the config's `outputDir` and
    is CLEANED at the start of every run — a later `-g` re-run silently
    deleted this task's first set of captures. Evidence now goes to
    `.evidence/phase-10-3/`, and the spec says why.)_
    
  • 10.4 Close the mobile items in the ledger or file follow-up TODOS with owners if emulator infra blocks (honest blocker note, not a silent skip). _(2026-08-13 s30: CLOSED. The emulator infra did NOT block — it was rebuilt and a real device pass ran — so this task is the disposition of the nine rows that pass produced, not a blocker note. Four are closed with calibrated locks; five are open with a named reason each, and none of the reasons is "ran out of time".

    **Closed, fixed and verified on the device:**
    
    - **EVE-VIS-221 (S1)** the assistant authenticates as the member — the
      401 that made every turn offline. Lock is calibrated.
    - **EVE-VIS-222 (S1)** the keyboard no longer covers the composer. Lock
      compares geometry against the `type=ime` inset frame, because
      `assertVisible` passes on a control under the IME.
    - **EVE-VIS-229 (S2)** the sheet no longer discloses a voice it does not
      have; disclosure and capability chips now read one value.
    - **EVE-VIS-223 (S1) — auto-scroll half.** The newest turn is brought
      into view; lock goes red without it.
    
    **Open, each for a stated reason:**
    
    - **EVE-VIS-223 (S1), second half — needs a product decision.** The
      conversation sits beneath six diagnostic panels; reordering them is an
      information-architecture change on panels carrying testIDs and e2e
      coverage. **This is the one item that still gates the phase**, and it
      is a decision, not a task.
    - **EVE-VIS-224 (S2) — a shared-constant decision.** Mobile's buffered
      turn inherits the STREAMING client's 30s as a TOTAL budget. The fix is
      to name and size a separate buffered budget (the measured max here was
      63.5s on the mandated cheap model), which is a judgement about how long
      a member should wait with no reply — and it interacts with 225.
    - **EVE-VIS-225 (S2) — depends on 224.** A pending affordance and an
      in-flight guard belong with whatever wait budget 224 settles on;
      shipping a spinner against a budget about to change would be guesswork.
    - **EVE-VIS-226 (S2) — needs a BFF route.** The app posts crash
      diagnostics to `/v1/mobile/diagnostics`, which the BFF does not serve
      (404 ×9 in one session). Adding a telemetry endpoint is BFF work with
      its own retention and shape questions, not a mobile edit.
    - **EVE-VIS-228 (S2) — needs a shared vocabulary change.** "Voice is
      unavailable (missing microphone)" names the wrong cause; the honest
      fix adds a downgrade reason to `libs/oshun/shell-assistant`'s
      interaction-modes, which reaches the web shell too.
    - **EVE-VIS-227 (S3) — half fixed.** The doubled slash is gone; the
      `CURR`/`ENT`/`ROOM` label break is room-chrome layout and goes with the
      Phase-12 sweep.
    
    **One honest instrument gap, carried forward rather than hidden:** the
    SelectionAsk TOUCH GESTURE cannot be driven in this harness. Under
    `isMobile: true`, both a double tap and a CDP-held long press return
    `0 chars / 0 chips` on two different pages — Chromium's touch-selection
    controller does not raise a selection headlessly. 10.3 verified the
    product half (chip appears, is thumb-tappable, stays in bounds) and left
    the gesture unclaimed. Closing it needs a real device or a browser whose
    touch-selection controller is available; it is NOT evidence that
    SelectionAsk is broken on touch, and must not be recorded as such.
    
    **Also recorded from the mobile pass, outside the ledger:** the member
    session did not survive a cold app start during this session (the
    launcher returned to `enabled=false` after a force-stop and relaunch,
    requiring re-signup). It was seen repeatedly but not isolated — it may be
    the dev-client reload rather than the app — so it is written here as an
    observation to chase, not filed as a defect on evidence that thin.)_
    

Phase 11 — Governance, cost, and safety surfaces#

  • 11.1 Budget exhaustion rendered humanely: drive a member session to the daily token budget (set a tiny budget env) — 429 copy, panel remains usable for deterministic answers, next-day reset message. Tenant budget variant on an admin session. _(2026-08-14 s31: CLOSED. Driving this for real rather than injecting it found that the ceiling did not exist, and then that nothing rendered it when it did. Five defects, four S1/S2, all fixed and calibrated-locked. Spec: e2e-inspect/phase-11-1-budget-exhaustion.spec.ts (5 cells green: cream and dusk × desktop and narrow, plus the tenant leg); evidence in e2e-inspect/.evidence/phase-11-1/.

    **Phase 2.7 had already fulfilled a 429 at the network boundary and
    proved the member never meets the number. That is why this task exists
    separately, and the difference is the whole finding:** there is no
    `page.route` anywhere in this file. The member spends real tokens against
    the real ledger, and the refusal — when it finally came — is the BFF's
    own.
    
    **EVE-VIS-230 (S1, FIXED) — the daily budget could not be reached at any
    budget.** `libs/shared/ai`'s OpenAI-compatible provider returned
    `usage: { inputTokens: 0, outputTokens: 0 }` from its STREAMING method
    behind the comment `// Not available in streaming`. Every assistant turn
    streams, so the ledger only ever added zero and
    `dailyOutputTokens >= budget` could not be true. **The server's own
    metrics are the evidence: `dailyTokenBudget: 1`, `totalTurns: 2662`,
    `outputTokens: 0` for BOTH bound models — 2,662 real turns across this
    initiative's whole history, nothing ledgered, and turn 2,663 still
    served.** The per-tenant ceiling was dead for the same reason and
    `GET /v1/assistant/metrics` reported zeros for every provider, which is a
    number that is wrong rather than absent. **The comment was also false, and
    that was measured rather than argued**: `stream_options: {include_usage:
    true}` on the bound model returns `{prompt_tokens: 10, completion_tokens:
    16, total_tokens: 26}` from the live API. Reading it also needed the guard
    moved — the spec's usage chunk carries `choices: []`, which
    `if (!choice) continue;` skipped. After the fix: one member turn reports
    `{inputTokens: 10938, outputTokens: 154}` and the next is **429 in 12
    ms**.
    
    **EVE-VIS-231 (S1, FIXED) — the refusal reached nobody.** With the ceiling
    alive, the member at it read: **"I lost my connection for a moment, so
    that one did not reach me — say it again and I will pick up right where we
    were."** Nothing was lost, and saying it again cannot work — the cap holds
    until the UTC day rolls over — so the member loops on that sentence with
    no cause and no end time. The humane copy existed on the server the whole
    time and no client rendered a word of it. Fixed at three levels: the 429
    now carries `resetsAt`, DERIVED from the same `todayUtc` key the usage map
    is written under (a client that computed "UTC midnight" itself would keep
    saying so after the ledger moved to any other window); the stream error
    carries `reason` + `resetsAt`; the panel says it ONCE, after which the
    deterministic engine's own copy stands, because that engine is the one
    answering now. The member now reads "That is all of my longer thinking for
    today. I can still find practices, open rooms and answer the quick things
    — those keep working as they were. They come back at midnight tonight."
    The admin drawer had the identical hole and got the operator-plain
    variant.
    
    **EVE-VIS-233 (S1, FIXED) — a conversation could break its own session,
    and this is the one that would have been missed.** Turn 3 of the "panel
    remains usable" leg answered "I'm having trouble reaching the assistant
    service right now". The deterministic engine writes the resolved intent's
    domain into `activeDomain` with an unchecked cast, and the resolver sends
    ordinary sentences to the two rooms V1.0 defers — **"Show me my saved
    practices" resolves to VERITAS on the word "saved"**. The checkpoint store
    then refuses the whole session (400
    `assistant_session_active_domain_invalid`) and every message after it
    fails. **This is EVE-VIS-030's symptom from a THIRD producer**: that row
    closed the personalization ranker and the context handoff in August; the
    intent path never consulted the allow-list the session already carries.
    The guard now restricts the active room to `authorizedDomains` — the exact
    set the validator checks — so the two cannot drift apart.
    
    **EVE-VIS-234 (S2, FIXED) — and what that sentence said instead.** With
    the session no longer breaking, the reply read "Sorry, I'm having trouble
    reaching **veritas** right now. Please try again in a moment", with a Try
    again chip carrying the raw intent name. A deferred room is not hard to
    reach, the slug is not member vocabulary, and the retry can never work.
    It now speaks the product's own sentence ("Veritas — evidence, claims, and
    grounded reading — is not part of V1.0. It opens in V1.2.") with no retry;
    shipped rooms are named by `displayName`.
    
    **EVE-VIS-232 (S3, FIXED)** "Confidence: 58%" was printed under the
    connection notice — a confidence score for a statement about
    connectivity, carried through from the reply whose text had been
    replaced. **EVE-VIS-235 (S3, OPEN)** the notice keeps Save-to-Library and
    Copy when it stands in for a reply and has neither when appended as a
    system line; the honest fix is to make it a system message in both
    branches, which touches the turn-completion flow, so it is owned by the
    Phase 11 boundary.
    
    **Tenant variant, on an admin session:** operators are not metered members
    (the route zeroes their plan cap explicitly) but the ORG ledger applies to
    them when the token carries `tid`. Driven at the HTTP boundary with two
    different operators in one org: the first is served, **the second is
    refused for the first one's spend** — which is what makes it a tenant
    ceiling rather than a per-person one — with `scope: 'tenant'`, the
    organisation sentence, and no slug.
    
    **Every lock was calibrated by mutation**: 4 of 5 provider tests, both
    deferred-intent tests, all three formatter tests and both admin-drawer
    tests go red against the code as it was, while their controls stay green.
    
    **Harness notes for the next session.** (1) The BFF must be booted with
    `OSHUN_ASSISTANT_DAILY_TOKEN_BUDGET=1` and the `TENANT_` variant — the
    spec reads them back off `/v1/assistant/metrics` and fails loudly rather
    than measuring an unreachable ceiling. (2) **The ledger cannot be read off
    the wire**: the panel cancels its stream reader the instant
    `turn.complete` lands, which errors the observer's tee branch, so the
    frame carrying `usage` is routinely lost — poll the server's metrics
    instead. (3) A reply bubble paints on the FIRST delta, so reading any
    server-side total the moment a bubble appears measures a turn that has not
    finished; that misread a working ledger as broken once. (4) "Show me my
    saved practices" now opens the immersive session player over the composer
    — a real do-tier action, and the reason 11.1's third turn asks about the
    sky instead.)_
    
  • 11.2 Kill switch (OSHUN_ASSISTANT_AGENT_ENABLED=0): member sees the deterministic engine seamlessly, no broken UI, no error toast storms. _(2026-08-14 s31: CLOSED. Two of the three clauses were already true — no broken UI, no toast storms, nothing to fix there. "Seamlessly" was not: with the agent deliberately off, the product told the member a connection had dropped. EVE-VIS-236 (S2), fixed and calibrated-locked. Spec: e2e-inspect/phase-11-2-kill-switch.spec.ts, 4 cells green; evidence in e2e-inspect/.evidence/phase-11-2/ (16 screenshots).

    **Driven under the real switch, not an injected 503.** Phase 2.7 fulfilled
    a single 503 and proved no jargon reaches the transcript; the kill switch
    is a different question, because it is not one failure — it is the state
    the product is in for as long as an operator leaves it there, so the
    thing to look at is what ACCUMULATES. Four turns per cell, two of them
    deliberately beyond the deterministic engine.
    
    **EVE-VIS-236 (S2, FIXED).** Turns 2 and 4 came back, in all four cells,
    with **"I lost my connection for a moment, so that one did not reach me —
    say it again and I will pick up right where we were."** Nothing had
    failed, and saying it again cannot change anything until an operator
    flips the flag back. **The panel's own code had already reasoned this out
    and guarded the wrong half**: its comment says that with streaming off
    "that route is the PRIMARY engine and nothing has gone wrong, so telling
    the member a connection dropped would be a fabrication" — true, and it
    only covers the CLIENT-side flag. The server-side switch leaves the
    client streaming, so it asks, gets the 503, and takes the failure path
    anyway. **And there was a test sitting right next to it that passed
    throughout** (`never claims a dropped connection when no agent turn was
    attempted`) because it mocked the client flag to false — the exact shape
    of "a test that derives an expectation the same wrong way the code does".
    The rule is now stated once: states an OPERATOR chose
    (`assistant_agent_not_configured`, `assistant_budget_exhausted`) are not
    outages, and only a turn that genuinely died in transit may report one.
    Under the switch the deterministic engine IS the product, so its own
    not-understood advice stands, chips intact.
    
    **What was already right, and is now measured rather than assumed:** no
    alarm accumulates over four turns (0 throughout), the composer stays
    live, no modal opens over the conversation, and both console lenses are
    clean in every cell. The alarm count carries a **planted-alarm positive
    control** inside the spec, because "zero alarms" and "a lens that sees
    nothing" read identically — the same reason 9.5's kanban check passed
    over an empty set.
    
    **One instrument fault, found by looking at what the instrument had
    counted.** Both narrow cells failed on "a modal opened over the
    conversation" while their screenshots showed a calm, correct transcript:
    at 390px the panel's own presentation is `role="dialog" aria-modal="true"`
    **on purpose** — that claim is what keeps Tab inside it (EVE-VIS-045).
    The lens now excludes the panel itself. The product was never wrong here.
    
    **Corroborates an open row rather than duplicating it:** EVE-VIS-087
    says the kill switch silently changes the product's CAPABILITIES, not
    just its engine. Measured here: "summarise everything we have discussed
    so far in detail" is an agent question, and under the switch it returns
    the rephrase copy. That row is owned by the Phase 4 boundary; nothing new
    was filed for it.
    
    **Harness note:** `.artifacts` is the config's `outputDir` and is CLEANED
    at the start of every run, so a single `-g` calibration re-run deleted
    three of four cells' captures. This spec now copies each screenshot into
    `.evidence/phase-11-2/` itself — EVE-VIS-104's class, third sighting.)_
    
  • 11.3 Refusal path: elicit a provider refusal; turn.error → fallback copy is calm and doesn't leak provider jargon. _(2026-08-14 s31: CLOSED. The refusal could not be elicited by asking, and when it was finally driven it turned out the path had never been reachable at all. EVE-VIS-237 (S1) and EVE-VIS-238 (S2), both fixed and calibrated-locked. Spec: e2e-inspect/phase-11-3-refusal-path.spec.ts (2 themes green); evidence in e2e-inspect/.evidence/phase-11-3/.

    **"Elicit a provider refusal" — measured first, and it does not happen.**
    Three disallowed prompts on the live API (`deepseek/deepseek-v4-flash-0731`
    via OpenRouter, price-sorted) all came back `finish_reason: "stop"` or
    `"length"` with the refusal written in TEXT: "I cannot provide
    instructions for making firearms, especially untraceable ones…". A model
    declining in prose is a normal completed turn, and rendering it as one is
    correct. So the API-level refusal is a different event, and it is the one
    the BFF built its honest surface for.
    
    **EVE-VIS-237 (S1, FIXED) — that surface was unreachable.** The
    OpenAI-compatible adapter mapped `content_filter` to **`end_turn`**, so a
    provider that refused was reported as having finished normally and the
    partial text a filter left behind was handed to the member as the answer.
    `runAssistantAgentTurn`'s `stopReason === 'refusal'` branch and the turns
    route's `turn.error` / `assistant_agent_refusal` ("Honest surface — no
    fabricated answer") could not fire for **any provider this product
    binds**. The same flattening was in `providers/xai.ts`; the two
    round-trip consumers are now honest too (`provider-bridge` maps `refusal`
    to its own `content_filter` instead of defaulting to `end_turn`, and
    `codex-server`'s OpenAI-compatible surface reports `content_filter`
    rather than `stop`). That the repo already treats `content_filter` as
    first-class in the agentic bridge is what makes this a flattening rather
    than a design choice.
    
    **How it was driven, and why not by faking the route.** The vendor is
    doubled at its HTTP boundary
    (`e2e-inspect/support/chat-vendor-double.ts`, BFF booted with
    `OPENROUTER_BASE_URL=http://127.0.0.1:4789/v1`) and everything else is the
    product: the real panel, the real turns route, the real agent runner, the
    real `OpenRouterProvider`. Fulfilling `/turns` with a hand-written
    `turn.error` would have skipped the adapter and the runner — the two
    pieces actually under test — and would have passed cleanly throughout the
    defect. **The spec's first leg is a control**: it asserts the double
    received a request, so the refusal leg cannot quietly measure a live
    model if the BFF is booted without the base URL.
    
    **EVE-VIS-238 (S2, FIXED) — and then the copy.** A refused turn took the
    same client path as every other stream failure, so the member read "I
    lost my connection for a moment … say it again and I will pick up right
    where we were." Nothing was interrupted, and repeating the request is the
    one thing least likely to change the answer. **Third instance of one
    class in this phase** — after the budget ceiling and the kill switch — so
    the rule is now stated once: `assistant_budget_exhausted`,
    `assistant_agent_not_configured` and `assistant_agent_refusal` are not
    outages, and only a turn that died in transit may report one. The member
    now reads "That one is outside what I will take on. Ask me another way,
    or ask me something else — I am still here." **The server's own sentence
    was deliberately not reused**: "The assistant model declined this
    request" is honest and addressed to an operator; "the assistant model" is
    machinery on a surface where Lilith is the only voice, and it is on the
    spec's jargon sweep for that reason.
    
    **Observation, not filed:** `src/agentic/autonomy-bindings/media-best-of-n.test.ts`
    is red on this branch (2 of 3). Verified pre-existing by stashing this
    task's changes and re-running — it arrived from main and belongs to
    whoever owns that binding.)_
    
  • 11.4 Safety supersede: a prompt that trips the Lilith safety layer renders the safety response with correct visual treatment (and never the raw model text). _(2026-08-14 s31: CLOSED, and this is the most consequential task in the phase. Two S1s, both in the crisis path, both live on every crisis reply. Spec: e2e-inspect/phase-11-4-safety-supersede.spec.ts (2 themes green); evidence in e2e-inspect/.evidence/phase-11-4/.

    **EVE-VIS-239 (S1, FIXED) — a member in suicidal crisis was reading the
    system's own tuning notes.** The reply ended: "…I am a contemplative
    teacher, not a clinician — this is out of my scope, but help is
    available. **Continue with grounded tone guidance.**" That sentence is an
    instruction addressed to the MODEL. `recommendations` carries two
    audiences at once — the crisis analyzer fills it with
    `plan.statementsToDeliver`, which `buildCrisisResponseText` joins into the
    sentence a member reads, while the BFF's scope checks were putting conduct
    rules in the same array — and `mergeLilithSafetyAnalyses` concatenates
    them. **It fired on 100% of crisis replies**, because the offending note
    sits on the SAFE branch, which is the default every crisis merge picks
    up. The scope checks now contribute a member-deliverable sentence or
    nothing.
    
    **EVE-VIS-240 (S1, FIXED) — the product told a member in crisis to "reach
    out to a crisis line" and did not give them the line.** The supersede
    resolves real regional help and sends it on `turn.complete` —
    `988 Suicide & Crisis Lifeline` with the number, `Crisis Text Line` with
    "HOME to 741741", `Emergency Services` with `911`, each with a priority —
    and **no client read that payload**. The panel's turn type did not declare
    it, so it was dropped. The most important content this product produces,
    held by the server and hidden from the person it was resolved for. It now
    renders as a bordered block inside the bubble, ordered by the catalog's
    own priority, each row a one-press `tel:` or link, and nothing is
    invented — a resource with no route renders as its name and stops.
    
    **EVE-VIS-241 (S3, FIXED)** "Confidence: 100%" under a crisis reply. A
    supersede resolved nothing; the meter is certainty about a question that
    was never answered. Four hardcoded sites, both routes.
    
    **"Never the raw model text" — the honest verdict is that it holds at the
    END of the turn and not during it (EVE-VIS-242, S2, OPEN).** Measured on
    the wire: deltas carry the model's words, `turn.complete` carries the
    safety sentence, and the bubble does settle correctly. But the member can
    read the disallowed text for the rest of the stream. Closing that window
    means running the catalog over a partial generation and stopping
    mid-flight, which is a policy change with a real false-positive cost — a
    half-sentence is not a reliable input to a crisis catalog — so it is
    flagged rather than patched.
    
    **The control that keeps leg 1 honest:** a BLOCKED turn must render NO
    resources. A resource row that appeared under every reply would satisfy
    "the resources are rendered" without meaning anything.
    
    **Instrument fault, and the lesson is the same one 11.1 learned in a
    different costume.** The post-hoc leg first reported the supersede as
    broken — the model text was in the panel 600 ms after the bubble appeared
    — while the screenshot taken moments later showed only the safety
    sentence. **A bubble paints on the FIRST delta, so "a reply appeared" is a
    mid-stream event**, and the post-hoc supersede is precisely a turn whose
    text changes at the end. The spec now waits for the newest reply to stop
    changing. A first version of that wait then read the GREETING as the
    answer in the dusk cell — stable, and the wrong bubble — so the settle
    condition requires the reply COUNT to have moved as well.
    
    **Pre-existing and untouched:** `src/adapters/arete-oshun-facade-contract.test.ts`
    fails 10 of 11 on this branch; verified by stashing this task's changes
    and re-running.)_
    
  • 11.5 Feedback thumbs: render, submit, persist (revisit shows state), copy; verify feedback stores label only (no content) via row check. _(2026-08-14 s31: CLOSED. Render, submit, copy and the row check were already right. "Persist (revisit shows state)" was not, and it failed for two independent reasons — EVE-VIS-243 and EVE-VIS-244, both fixed and separately calibrated. Spec: e2e-inspect/phase-11-5-feedback-thumbs.spec.ts (2 themes green); evidence in e2e-inspect/.evidence/phase-11-5/.

    **The revisit is a REAL navigation** (nyx → tara → nyx), not a component
    remount, because that is what actually happens to a member and it is what
    makes the failure non-obvious: `ShellLayout` is rendered per page, so
    every client-side navigation unmounts the panel.
    
    **EVE-VIS-243 (S2, FIXED) — and the first cause hid behind the second.**
    `sanitizeAssistantMessages` rebuilds each restored message field by field,
    and **`turnId` was not among the fields** — the thumbs render only for a
    message that has one, so a restored transcript could not be given
    feedback AT ALL, not merely lose its highlight. `toolNote` was dropped the
    same way, taking the "Checked your recommended practices" provenance line
    with it (EVE-VIS-036's whole point). Then, even with the id back, the
    verdict was component state and the store was **write-only from the
    member's side**: one durable label per (userId, turnId), and no way for a
    client to ask for it. Fixed with a new
    `GET /v1/assistant/sessions/:sessionId/feedback`, scoped to the member AND
    the session, returning `{turnId: verdict}` and nothing else — restored
    from the SERVER rather than the snapshot, so the same member on another
    device sees the same verdicts, and merged UNDER anything pressed since so
    a slow read cannot undo a fresh press.
    
    **EVE-VIS-244 (S2, FIXED) — the fix did not work on its first version,
    and how it failed is the finding.** The read-back reads
    `tryGetApiAuthToken()`, the auth context's synchronous mirror, which is
    empty until `/api/auth/session` answers; fired on mount the request went
    out bare — **5 of 5 at 401 in the BFF's own log** — and the thumb came
    back un-pressed exactly as before. **Third instance of one class**, after
    EVE-VIS-184 (audit board) and EVE-VIS-214 (graph explorer), and the same
    remedy: wait for `authStatus === 'authenticated'`.
    
    **The row check passes, and is asserted two ways.** The member's own words
    ("sky") are absent from the row, AND the row carries no key outside the
    label vocabulary — so a future `note`/`prompt`/`text` field fails the
    check without anyone having to remember to look for it. The store's
    `note` seam exists but the panel never sends one, which the rows confirm.
    
    **Submission is witnessed by the SERVER's export, not by the button's own
    styling** — an optimistic UI would change that either way.
    
    **One harness lesson:** the dusk cell first died on `locator.click:
    Timeout 20000ms` with the thumb below the fold mid-scroll. A control that
    renders and cannot be pressed is a FINDING, not a timeout (EVE-VIS-026,
    EVE-VIS-196), so the spec now names what is on top of it before clicking
    — and the answer here was "nothing, it just needed scrolling", which is
    the honest verdict rather than a filed defect.)_
    
  • 11.6 Metrics/telemetry sanity behind the scenes: after the above, GET /v1/assistant/metrics totals move; telemetry sink rows are shaped (no free text) — spot-check via psql. _(2026-08-14 s31: CLOSED, and this is the one task in the phase that found nothing wrong — because 11.1 had already fixed the thing that was. Spec: e2e-inspect/phase-11-6-metrics-telemetry.spec.ts, green.

    **"Totals move" is a stronger assertion than it looks, and that is the
    whole reason this task follows 11.1.** Before EVE-VIS-230 the token
    totals were structurally incapable of moving — the streaming adapter
    reported zero usage on every turn, so `/v1/assistant/metrics` read
    `inputTokens: 0` / `outputTokens: 0` across 2,662 recorded turns. **A
    version of this check that only watched `totalTurns` would have passed
    throughout the defect**, which is why the spec asserts the token sums and
    names the row it is guarding.
    
    **"No free text" is checked with the member's own words as the needle.**
    The sink redacts by SHAPE — numbers, booleans, strings ≤80 characters —
    and a member's message is usually shorter than that ("I want to kill
    myself" is 21), so the rule cannot be verified by reading it. The spec
    sends a turn carrying a nonce nobody else would say and asserts it
    appears nowhere in telemetry or metrics. It also walks every recent row's
    `data` and fails on any non-scalar or over-length value, and asserts the
    row population is non-empty first — a shape check over zero rows passes
    for the wrong reason.
    
    **The psql spot-check the task asks for, run against the live database:**
    
    - `assistant-telemetry` — `{counters: {turn.started: 249,
      action.failed: 10, turn.completed: 235, action.executed: 103,
      session.created: 4474, intent.classified: 249, navigation.triggered:
      5}, totalEvents: 5325, firstEventAt: 2026-08-05, lastEventAt}`. Nine
      days of events, 309 bytes, **no rows at all** — the durable record is
      counters and timestamps, exactly as the module's header claims.
    - `assistant-feedback` — the distinct key set across EVERY persisted
      entry is `userId, sessionId, turnId, verdict, providerId, model,
      createdAt, updatedAt`. **No `note`, no text, nothing from a
      conversation** — which durably confirms 11.5's row check across the
      whole store rather than for the single row that task created.
    - `assistant-turn-metrics` — `aggregates, dailyUsage, dailyTenantUsage,
      totalTurns, schemaVersion`. Counters only.
    
    No ledger rows: everything this task inspects is right, and it is right
    because the ledger it depends on was repaired first.)_
    
  • 11.7 Disclosure & memory consent UI: the disclosure chips' information architecture (what does "Memory: session" MEAN to a member — copy review), memory opt-out honored visibly. _(2026-08-14 s31: CLOSED with no defects — the second clean task in this phase, and it is clean end to end: profile switch → shell state → context handoff → the words on the panel. Spec: e2e-inspect/phase-11-7-disclosure-and-memory.spec.ts (2 themes green); evidence in e2e-inspect/.evidence/phase-11-7/.

    **The copy review's answer: the raw chip is not what a member reads.**
    `Memory: session` does exist — the literal enum, rendered as a chip — but
    it lives inside the context strip EVE-VIS-015 deliberately folded away as
    builder telemetry, and the spec asserts it is NOT visible unexpanded. What
    a member reads instead is prose: "The assistant can use the current
    conversation while this session is open", "This conversation only". So the
    task's question ("what does Memory: session MEAN to a member") has a good
    answer already, and the check now holds the product to BOTH halves — the
    scope must be named in readable prose AND explained, because "Session
    memory" alone does not tell anyone whether their words are kept.
    
    **Opt-out driven through the member's own control**, not by writing state:
    the "Off the record" switch on `/profile/memory`. After it, the panel's
    operating scope is `off` and the "WHERE YOU ARE" card's chip reads
    **"Remembers nothing"** where it read "Remembers this conversation". The
    spec also asserts the pre-opt-out promise is GONE rather than merely
    joined by a new sentence — a disclosure that adds "memory is off" beneath
    "remembers this conversation until you end the session" would satisfy a
    naive check and contradict itself on screen.
    
    **Calibrated**: making the shell ignore the pause flag turns the cell red
    on "the member opted out and the panel is still operating on session
    memory".
    
    **Measured with `checkVisibility()` throughout**, because a closed
    `<details>` still has measurable children (Chrome hides them with
    `content-visibility`) — the trap that once had a contrast walk reading
    four chips nobody can see, and here the difference between "the panel
    discloses this" and "the panel could disclose this if you went looking".)_
    

Phase 12 — Whole-estate visual sweep (assistant-adjacent pages)#

  • 12.1 /welcome + signup + onboarding (all 10 steps + skip dialog): five lenses, both themes, three viewports (EVE-VIS-008 regression included). _(2026-08-14 s32: four defects, two of them S1, and the pair that matters is one sentence: the product a visitor is SOLD is a six-room product, and the one they GET has four. Spec: e2e-inspect/phase-12-1-welcome-signup-onboarding.spec.ts — 6 cells (2 themes × desktop/390/200%-zoom), each one a real signup through the real form and then all ten steps walked with clicks, because three steps gate their own Continue and a ?step= jump would never learn that. Evidence in e2e-inspect/.evidence/phase-12-1/.

    **EVE-VIS-245 (S1)** — the release-scope module says no console may
    "advertise, link to, count, or compose" Veritas or Metis before V1.2, and
    the acquisition funnel did all four: 15 visible mentions, three wrong
    counts on one page (`6` on a metric tile, "Six domains, one practice"
    over a four-card grid, "across all six" — on a page whose hero says
    "Four rooms"), and two of four public conversion cards carrying
    `redirect=/domains/veritas|metis` through account creation, under a
    subtitle promising every public route survives sign-up. Following one as
    a member lands on "Veritas arrives in V1.2". **The cut had been applied
    to the LISTS and never to the copy, the counts, the links or the
    options** — visible on one screen in onboarding, where `GoalCard` looks
    each room up in the scoped `DOMAINS` and renders nothing when it misses,
    so two goal cards showed an EMPTY chip row while their own prose still
    said "Use Veritas…".
    
    **EVE-VIS-246 (S1)** — at 390px the service-worker notice is a
    full-width ~800px sheet and `/welcome` has no shell chrome for it to sit
    above, so it buried the display-name field, the email field and all three
    auth toggles: no account could be created while it was up. EVE-VIS-008
    recurring at a width where "move it to the other corner" has no other
    corner. It now opens as a 108px bar with `Details`.
    
    **EVE-VIS-247 (S2)** `/profile` personalization offered eight controls
    pointed at deferred rooms; **EVE-VIS-248 (S3)** `/welcome/domains`
    rendered a skip link to a `#main-content` that does not exist.
    
    **The finding that explains the whole task: four tests in
    `OnboardingWizard.test.tsx` had been RED on this branch since the release
    cut, all four asserting the product offers these rooms — and the half of
    the same suite that would have caught the copy was GREEN.** A partly-red
    suite reads as "the cut's expected fallout" rather than as a report that
    the component was never cut.
    
    **Two instrument faults, both of which produced a false defect first.**
    (a) The occlusion leg centred `#auth-entry` and reported "Create your
    account" buried — at a scroll position no link in the product produces;
    it now scrolls the card's LAST control to the top of the viewport, which
    is the reachability question. (b) A body-padding reservation was written
    for that false defect, and **measuring it showed it changed nothing** —
    identical covered set with and without at maximum scroll — so it was
    removed rather than shipped with a plausible comment
    (`probe-12-1-toast-bottom-reservation.spec.ts` keeps the comparison).
    Every lock calibrated by mutation: 4 for 245, 3 for 246, 1 each for 247
    and 248.)_
    
  • 12.2 Home (all dayparts — morning/midday/evening variants): the daypart rail, catch-up cards (date formats!), tour anchor targets. _(2026-08-14 s32: six defects, three S1, and the worst of them is one sentence: a member who signed up seconds ago was shown a practice they had never had. Spec: e2e-inspect/phase-12-2-home-dayparts.spec.ts — 6 cells × 3 dayparts = 18 inspections, the clock driven with Playwright's clock API because useClientHour() reads the real hour in a post-mount effect and a run at 13:00 local can otherwise only ever see midday. Evidence in e2e-inspect/.evidence/phase-12-2/.

    **EVE-VIS-253 (S1)** — four components default to hand-written fixtures
    when given no data, and Home (and `ShellLayout`) mount them with none:
    "Meditation Streak 12 days", "Daily Habits 4/6 completed", "Claims
    Verified 23 this week", "Weekly Minutes 145 +12%", a "Mindful Morning"
    routine already 2/5 done today, and a routine banner on EVERY page for a
    session never started — still reading "morning" at 21:10. EVE-VIS-010's
    shape, four more times. **EVE-VIS-254 (S1)** — and it had a second
    source: the BFF seeds a new member's profile through the DEMO factory, so
    the server itself said seven-day streak, five active domains. Three
    client-side zeroings did not move the number, which is how the server was
    found.
    
    **EVE-VIS-249 (S1)** — Home carried more deferred-room surface than the
    marketing site: a whole "Grounded evidence" pane for Veritas and a LIVE
    Metis assessment card, because the multi-panel branch hand-mounted them
    around the release-scoped `HomeCompanionSections`. Plus **250** (a raw
    ISO instant in a notification body, from four BFF `meta` strings), **251**
    (a frozen fixture date rendered as "Updated 143d ago"), **252** (search
    filter chips for rooms that do not exist).
    
    **The method that produced all six: read the RENDERED page, at a fixed
    hour, with an account created in the same run.** A screenshot shows the
    numbers; only a fresh account proves they are not the member's.
    
    **Two instrument faults, both caught before they became findings.** The
    anchor check calls `scrollIntoView` — and the shell sets `scroll-behavior:
    smooth` (EVE-VIS-109), so the rect read on the next line is the PRE-scroll
    one: it reported the daypart rail and the domain switcher as "covered" at
    an offset they were never at. And leaving the page where that scroll ended
    had the overlap lens report the sticky breadcrumb over a rail step, which
    is what a sticky header does. `behavior: 'instant'`, and scroll home
    before the lenses run.
    
    **The rail is checked by its ARC, not its presence**, and the cell asserts
    the three arcs DIFFER — three identical readings would otherwise pass
    three times if the clock never moved.)_
    
  • 12.3 Each of the 6 domain rooms' shell chrome as the assistant sees/anchors it (anchor targets visible and stable). _(2026-08-14 s32: the registry promised the assistant an anchor that existed on one page. Spec: e2e-inspect/phase-12-3-room-shell-chrome.spec.ts — 6 cells, each walking the four rooms V1.0 opens plus the two it defers. Evidence in e2e-inspect/.evidence/phase-12-3/.

    **EVE-VIS-255 (S2)** — `shell.domain-switcher` is registered
    `availability: 'always'`, `routePrefix: '*'`, and those summaries go
    verbatim into the agent's system prompt — but the attribute was stamped
    only on a HOME-ONLY panel. On every room page it was absent, so
    `shell-orientation`'s switcher step and any `highlight` aimed there found
    nothing. The shell's sidebar has had the room list on every page all
    along; it just had no name. Fixed by naming it there, through a FLAG
    rather than a caller-supplied id, with the Home panel keeping the anchor
    only where the sidebar is hidden — exactly one element per page, because
    a tour resolving by id spotlights the first match.
    
    **A wrong turn worth recording: I changed `assistant.composer` from
    `always` to `conditional` and it was wrong.** It reads like the same
    defect — the composer is absent until the panel opens — but `host:
    'assistant-surface'` is the field that carries that, the tour player
    mounts the panel for exactly those steps, and `tour-plans.spec.ts`
    requires every curated step's anchor to be `always`. The change would
    have broken `assistant-orientation`, the tour that field exists to make
    work. Reverted before commit.
    
    **Two rows left OPEN with their remedies named, not swallowed.**
    **EVE-VIS-256 (S3)**: below 769px there is no room switcher in the chrome
    at all — the honest fixes are product decisions (give mobile a switcher,
    or make the step viewport-aware), and stamping the anchor on the Explore
    tab would point a member at a different control and call it the switcher.
    **EVE-VIS-257 (S2)**: at 200% zoom the floating utility dock covers each
    room's `h1` — "Meditation & m…" cut off mid-word — because the shell
    publishes `--shell-bottom-chrome` and the content area does not consume
    it. The cell allows exactly that one overlap and then **asserts it is
    still there**, so the row cannot rot.
    
    **The check that matters most is STABILITY**, not presence: the cell
    requires every room to agree about which chrome it has, because a
    switcher on Tara and none on Nyx is a tour that works until it doesn't.)_
    
  • 12.4 /assistant entry route + AssistantEntryRoute behaviors. _(2026-08-14 s32: CLOSED with no defects — the first clean task in phase 12. Spec: e2e-inspect/phase-12-4-assistant-entry-route.spec.ts (6 cells); evidence in e2e-inspect/.evidence/phase-12-4/.

    Driven, not eyeballed: the panel opens by itself at all three widths (a
    real question, because `openAssistantInvocation` REFUSES below the
    invocation point's `minViewportPx` and warns to the console rather than
    throwing — a route that failed to open on a phone would just look like a
    page with a button on it); a `?prompt=` link asks the LINK's question and
    never the default; a 800-char link is truncated at 600 with an ellipsis;
    and after Escape the page's own button brings the panel back, which is
    the only way back without a reload since the auto-open fires once per
    mount.
    
    **Two corrections to my own assumptions, both caught by driving it.**
    (a) I wrote the spec expecting the prompt to be SEEDED in the composer
    and not sent. It is sent — and that is the design: a deep link that
    carries a question asks it, and EVE-VIS-021 was never about the sending,
    it was about the route IGNORING `?prompt=` and sending a different
    question. The check is therefore about CONTENT: every member turn is the
    link's question, and the default must NOT appear. (b) The first version
    read the BARE leg's default prompt out of the deep-link leg's transcript
    and called it a substitution — the panel restores its transcript from a
    client snapshot that survives an account swap, so each leg now clears the
    origin's storage before reseeding.
    
    **A dev artifact deliberately NOT reported as a defect:** one arrival
    really does reach the server twice, with two distinct session ids —
    `reactStrictMode: true` in `next.config.mjs` double-invokes the effect.
    The spec asserts turn CONTENT and never a turn COUNT, because a count
    assertion here would be asserting the absence of StrictMode.)_
    
  • 12.5 Admin: operations dashboard + the workspaces the Copilot references, so its citations point at real, polished UI. _(2026-08-14 s33: CLOSED. Sweep e2e-inspect/recon-12-5b-admin-workspace-sweep.spec.ts over all 20 workspaces in OSHUN_ADMIN_WORKSPACE_MODEL, not the four the s32 recon reached; lock e2e-inspect/phase-12-5-admin-operations.spec.ts (6 cells); evidence under e2e-inspect/.evidence/.

    **Four defects, and the biggest one was not the heading.**
    **EVE-VIS-259 (S1)**: between 721px and ~1550px — every laptop width,
    including this harness's 1470 reference and 200% zoom's 735 — the admin
    header's control row had `flex-wrap` only inside its `max-width: 720px`
    phone query, so above 720 it could only relieve pressure by SHRINKING.
    `.operator` had `min-width: 0` and `overflow-wrap: anywhere`, which
    together have no floor: the operator's own name collapsed to **0px wide
    and 33 lines**, one letter per line, and that strip became the header's
    height (**586px**), so most of the dashboard's first screen was empty
    header. Sign out was pushed off the right edge and the document scrolled
    sideways on 19 of 20 workspaces. **EVE-VIS-258** widened from 4 pages to
    **18 of 20 with no `h1` at all** and is now the SHELL's job, derived from
    the same IA entry the breadcrumb and the Copilot's context band use.
    **EVE-VIS-260 (S2)**: the dashboard's 20 workspace cards — the ones
    carrying the status badge, the open/urgent counts and the SLA — had no
    anchor, no button, no handler, and a click left the operator where they
    were. **EVE-VIS-261 (S3)**: one page had no `metadata` and its tab read
    plain "Oshun Admin".
    
    **What passed, measured rather than assumed:** the Copilot names the
    workspace it is actually in on all five sampled routes (`guard: allowed`,
    context band equal to the IA label), the "Live as of" timestamp is real,
    the counts come from the admin BFF, and after the fixes all 20 workspaces
    report zero console errors, zero failed requests, zero occluded controls
    and zero contrast failures.
    
    **Two things about the admin app that change how it must be inspected.**
    (a) **There is no light theme** — `globals.css` declares
    `color-scheme: dark` and no `prefers-color-scheme` block exists under
    `src/`, so "both themes" is measured here as **both DENSITIES**
    (`data-admin-density`, comfortable | compact); cream/dusk would be the
    same pixels twice. (b) The dev server **restarts itself** when webpack's
    heap crosses Next's threshold, and a 20-route sweep triggers it reliably —
    three `/models`, `/rights` and `/review` "timeouts" were that, so every
    navigation in the lock goes through a retrying helper.
    
    **A harness fault that had been manufacturing defects all along, found
    here and fixed in `support/lenses.ts`:** the sweep reported a React
    hydration mismatch on `/analytics` and `/rights`, and the only prop React
    named was `style={{caret-color:"transparent"}}` on every field — which
    **Playwright** writes before a screenshot (`screenshotter.js:72`) and the
    app never writes at all. Measured 5 runs per arm: never-screenshotted
    **0/5**, `caret:'hide'` (the default) **4/5**, `caret:'initial'` **0/5**.
    `inspect()` now screenshots with `caret: 'initial'`, and the whole admin
    estate's console lens came back clean afterwards.
    
    **And the calibration needed calibrating.** The cell that re-imposes the
    pre-fix header CSS reproduced the crush (0px, 33 lines) but NOT the
    overflow, because it left the shipped `justify-content: flex-end` in
    place and the crushed row then overflowed LEFTWARD, where
    `document.scrollWidth` cannot see it. Overflow direction is a property of
    the alignment. With `flex-start` restored the revert reproduces the
    original reading exactly — 60px at 1470, header 586px, Sign out
    unreachable — which is what makes the width cell's green mean
    something.)_
    
  • 12.6 Reduced-motion + 200% zoom + keyboard-only traversal across the panel, tours, boards, explorer — one full pass each. _(2026-08-14 s33: CLOSED with no product defects — the second clean task in phase 12, and the reason to trust it is that four candidates were measured and WITHDRAWN rather than a bar being lowered. Recon e2e-inspect/recon-12-6-motion-zoom-keyboard.spec.ts + recon-12-6b-containment-escape.spec.ts; lock e2e-inspect/phase-12-6-motion-zoom-keyboard.spec.ts (15 cells); lenses in e2e-inspect/support/traversal.ts.

    **What passed.** Reduced motion: zero animating elements on all four
    surfaces, against a no-preference CONTROL of 40 — the control is in the
    lock, because "zero under reduce" is also what a broken lens returns.
    200% zoom (735×450): no horizontal overflow, no covered heading, and
    every covered control reachable after a scroll. Keyboard: no trap, no
    stranded stop, a visible focus indicator on every stop, on all four.
    The panel at 390px **really** contains focus — `aria-modal="true"` and a
    planted decoy button appended outside it as the last child of `body` is
    never reached, which is the only version of that check that means
    anything (DOM order alone would satisfy it otherwise). Escape closes the
    layer that owns it: the panel returns focus to its trigger, and a tour
    ends the TOUR and hands the conversation back rather than closing both.
    
    **Four withdrawn candidates, each for a stated reason.** (a) `Start the
    audit` under an `svg` — covered at rest ONLY; clear after a scroll and it
    answers a trial click. An occlusion reading is a reading about a scroll
    position. (b) `Scroll to top` "unreachable" on board and explorer — the
    control is `disabled`, `aria-hidden` and `tabIndex={-1}` until the page
    has scrolled 500px; a refused click on a deliberately inert control is
    the control working. (c) `Save to Library` behind the composer — not
    rendered at all for an ungrounded reply; `Copy this reply` is reachable.
    (d) The tour leaving 7 shell controls focusable behind its 55% scrim —
    the tour is non-modal BY CONSTRUCTION (`role="presentation"`,
    `pointer-events: none`), the step contract is `{anchorId, narration,
    route?}` with no action step, focus starts inside the card, its three
    controls are the first three stops and one Shift+Tab returns from the
    fourth. A focus trap would also take the composer away during
    `assistant-orientation`'s last step, which is a step ABOUT the composer.
    What the lock DOES assert about the tour is the thing a coach mark may
    never do: cover the element it is pointing at.
    
    **Three instrument faults, all in the first run, all fixed before any
    reading was believed.** (1) The traversal reported **0 stops** on a
    surface with 148 controls, because it treated the first
    `activeElement === body` as the end of the tab ring — Chromium's ring
    crosses the document legitimately. (2) It reported **no focus indicator**
    on the entire shell primary nav, because it looked for an outline or a
    ring-shaped shadow; those links mark focus by changing background and
    text colour. The lens now measures a DIFFERENCE between the focused and
    resting computed styles, which is what "you can see where you are"
    actually means. (3) It reported **27 of 59 stops outside the viewport**,
    because focusing an off-screen control under the shell's
    `scroll-behavior: smooth` starts an animation and the rect on the next
    line is the pre-scroll one — EVE-VIS-109's trap again. Also: the zoom
    lens reported the assistant panel "covering" the page behind it, which is
    what a full-height overlay at 735px is for.
    
    **What is NOT covered here:** operating the board and the explorer
    end-to-end by keyboard (starting an audit, driving the explore lens) —
    that is 15.1's happy-path spine, not a traversal.)_
    
  • 12.7 Contrast measurements (not vibes): sample the actual rendered colors of chip text, timestamps, tour narration, board mono labels in both themes; every text node ≥ WCAG AA. Fix tokens, not instances, where failures cluster. _(2026-08-14 s33: CLOSED. Recon e2e-inspect/recon-12-7-contrast.spec.ts + probe-12-7-strip-eyebrow.spec.ts; lock e2e-inspect/phase-12-7-contrast.spec.ts (12 cells). Every visible text node under the panel, a running tour, the audit board and the graph explorer, in cream and dusk, composited down to the real backdrop and judged against ITS OWN AA bar — 4.5:1, or 3:1 for large text, which measureTextContrast now computes from the rendered font size and weight instead of leaving to the caller.

    **Two defects, and the interesting one is a family.** **EVE-VIS-262**:
    `DOMAIN_ACCENTS` mixes theme-scoped `var()` tokens with hex literals and
    the context strip paints the "Where you are" eyebrow in that value. In
    dusk: Home and Tara **2.26:1**, Nyx **1.24:1** — the accent is very nearly
    the dusk paper — while Arete (8.36) and Nisaba (7.11) passed, because
    theirs are tokens. Cream was fine everywhere, which is why it had never
    been seen. **EVE-VIS-263**: the "Interrupt response" button carried a DUSK
    ink as a literal and read **1.17:1** on cream — the control a member
    reaches for when the reply is going wrong, invisible on the light theme.
    
    **The fix is a rule, not three colours:** an accent is usable as text only
    if it is a `var()`. The panel's fallback had been returning
    `domainAccent`, so Home — in neither accent map — kept handing back the
    same rust literal even after Tara and Nyx were fixed. Worst room now
    **5.97:1**, from 1.24.
    
    **The lock states the invariant rather than the instances:** a cell reads
    the eyebrow's rendered colour in cream and in dusk and requires them to
    DIFFER. A hex literal cannot satisfy that in any room — which is the check
    that catches the family, instead of the one instance someone happened to
    open the panel in. A CALIBRATION cell repaints the eyebrow a hair off its
    own background and requires the walk to report it under 1.5:1.
    
    **The reading that was nearly certified: an empty surface.** The recon's
    first run reported the explorer clean at **five text nodes** and the board
    clean at **thirteen** — the explorer was showing "Builder scope required"
    (its two endpoints answer 403 for any member session) and the board was a
    fresh account's "no audit run is active". With the operator credential on
    the explorer's endpoints and 7.2's seeded member on the board, the same
    walk measures **246** and **1,154** nodes, and both are still clean. A
    contrast pass over a gate page is a pass over nothing, so the lock asserts
    a MINIMUM node count per surface.)_
    

Phase 13 — Copy editorial closure#

  • 13.1 Re-run the Phase 3.1 string inventory after all fixes; diff; every string has been either approved or rewritten (tracked in the inventory file — a string with no verdict blocks this task). _(CLOSED 2026-08-15 s34: 2,230 of 2,230 judged, 0 unjudged, and every verdict carries a note — keep 1,650 · not-member-facing 479 · rewrite 101. The blocking condition this task names is a string with no verdict; there are none. Of the 101 rewrites, 12 were applied in this pass, 57 are handed to 13.2 (the terminology sweep owns them by name), and the rest are recorded with their new wording for 13.2/13.3 — several against copy that no longer renders at all, kept so the sentence is right the day its adapter lands. Five defects came out of the ruling, three of them S1. Original s33 history below.

    (2026-08-14 s33: STARTED, NOT closed — the box stayed `[ ]` because **8 of
    2,230 strings had a verdict** and the task's own rule is that a string
    with no verdict blocks it.
    
    **Done: the re-run and the diff.** The miner had been broken since s30 —
    `assistant-markdown.tsx` moved to `libs/oshun/shell-assistant/src/markdown.tsx`
    in task 9.3, the SCOPE list did not follow it, and every run since has
    exited 1 with `missing scope entry`. That is how a member-facing surface
    leaves an inventory without anyone deciding it should. Scope repaired;
    the miner is green. **Diff against the 3.1 baseline: 2,094 → 2,230, +227
    added, −91 removed** — added mostly in `engine replies` (54), `member
    panel` (44), `audit board` (29) and `confirm cards & tools` (28).
    
    **Two defects came out of the FIRST 60 strings**, which is the argument
    for finishing the pass. **EVE-VIS-264 (S1)**: reading the "member dock"
    surface's labels sent me to a screenshot of Home, where the quick-actions
    rail was still rendering a `veritas` card and a `metis` card with live
    `/domains/…` hrefs under a paragraph counting "six core shell journeys" —
    on the surface 12.2 had closed. Fixed and locked. **EVE-VIS-265 (S1)**:
    the rendered sweep that confirmed the fix found the same thing on EIGHT of
    the nine member surfaces — Explore, Activity, Library, Profile and all
    four room pages, five of them with live links into the deferred rooms.
    Ledgered OPEN with a per-route map and a lock that allows exactly the
    known-open set and requires it to still be failing.
    
    **The workflow for the remaining 2,222** is
    `tools/eve-polish/string-inventory-batch.mjs` (print the next unjudged
    batch, optionally one surface at a time) and
    `tools/eve-polish/string-inventory-verdict.mjs` (merge
    `id<TAB>verdict[<TAB>note]` into the sidecar). Both are committed rather
    than left in `/tmp`, which is where they were written and would not have
    survived the session that needs them. Verdicts: `keep` · `rewrite` · `remove` · `not-member-facing`.
    Judged so far: root launcher, persistent context strip, tour spotlight,
    selection ask. Two patterns already worth reusing — thrown-Error messages
    in the client libs are `not-member-facing` because the panel renders its
    OWN copy for a failed turn (measured: "I'm having trouble reaching the
    assistant service right now"), and a control whose visible label says
    "Ask Lilith" while its `aria-label` says "the assistant" is one control
    with two names, which is `rewrite` and 13.2's business.
    
    **EVE-VIS-265 is CLOSED — all nine member surfaces are silent about both
    rooms, 0 mentions and 0 links, measured on the rendered pages.** It went
    12 links → 0 first, then 34 mentions → 0.** Four hand-written room lists became release-scoped —
    the room shortcut rail on every room page, the cross-domain
    recommendations (filtered at the RENDER boundary so it holds for the
    server payload too, not only the fixture), Explore's five catalogues, and
    Explore's "Open a room directly" grid, which iterated the keys of
    `SEARCH_DOMAIN_LABELS`, a map that correctly still names every room. A
    curated item is dropped when its ACTIONS lead into a deferred room, not
    only when it belongs to one. Home, Tara and Nyx are completely clean.
    What remains is COPY on six surfaces, and it is prose, not structure.
    
    **Three test files were holding the defect in place** — cells titled
    "keeps Metis as a first-class learn lane", "filters curated discovery into
    a first-class Metis learn lane", and a launch-order assertion naming all
    six rooms. Each was rewritten to assert the scoped behaviour AND paired
    with an in-scope room, so none of them can pass on an empty catalogue.
    Separately, 46 failures elsewhere under `src/components/__tests__` were
    checked against the stashed tree and are **pre-existing** — identical
    counts before and after.
    
    **2026-08-14 s34: 57 of 2,230 judged**, and the third defect in a row came
    out of the ruling rather than the reading. 49 of the mined strings still
    name a deferred room; ruling on them meant asking, for each, whether it
    can reach a member — and for the ones in `action-router.ts` and
    `response-formatter.ts` the answer was yes. **EVE-VIS-266 (S1)**: the
    deterministic engine — what the panel falls back to whenever an agent
    turn fails, and the ONLY engine left past a member's daily token ceiling —
    had no release cut anywhere between the intent resolver and the member.
    Eight ordinary sentences ("open tutoring", "recommend me a course",
    "what's trending") were answered with Metis and Veritas content, cards,
    and `navigateTo` into rooms V1.0 does not ship; three of them consult no
    adapter at all, so no backend outage could ever have masked them. Fixed at
    two seams and locked at three layers, every lock calibrated red.
    **The most useful thing it taught:** seven cells in
    `assistant-session-release-scope.spec.ts` already covered deferred-room
    sentences and every one stayed green throughout, because they assert the
    SESSION survives and never read `response.text`. A lock can be about the
    right sentence and still never look at it.
    
    **The "engine replies" surface is fully ruled — 480 of 2,230 judged —
    and it produced two more S1s from the same reading.** The tell was a
    vocabulary that does not belong to a web release: "sync my watch", "Tray
    state", "Toggle mini player", "keyboard shortcuts configured".
    **EVE-VIS-267 (S1)**: `AssistantDomainAdapters` carries six room adapters
    and nothing else, so the `achievement.*`, `wearable.*` and `desktop.*`
    families had no one to call — and rather than say so, `callAdapter`
    answered them itself with a hand-written echo of its own request, which
    the formatter read for fields that were never in it. A V1.0 web member
    was told **"Watch data synced! Updated 0 complications and 0 reminders."**
    at **0.96 confidence**, **"Focus mode is now on. Notifications are
    paused."** with nothing paused, and **"You're Level 1 with 0 points."** as
    a fact about their account. **EVE-VIS-268 (S1)**: the generic `default:`
    branch replied **"Done!"** whenever an intent resolved to zero actions —
    so "add this meditation to my favorites" saved nothing and said it had.
    Both fixed by failing loud, both locked and calibrated.
    
    **The lesson for the rest of this pass:** a string list is a defect
    finder, and the finding is rarely the string. "Done!" reads as harmless
    in a column of 2,230; in context it is a claim. Read the surface's
    vocabulary as a whole and ask what release it belongs to.
    
    **The remaining surfaces produced three more, all from the same reading.**
    **EVE-VIS-269 (S3)**: the confirm card that writes a Tara favourite spelled
    it the British way — the only one in the product's rendered copy, and
    EVE-VIS-098's own comment quotes the string it replaced as American.
    **EVE-VIS-271 (S3)**: the audit board contradicted itself without
    scrolling — "every **catalogued** flow" above "removed from the
    **catalog**" — and the same drift ran through the audit page, the persona
    page and Lilith's tool notes. **EVE-VIS-272 (S2)**: the grounding badge's
    `abstained` description read *"The system declined to answer and should
    render fallback copy only"* — an instruction to the RENDERER, on a field
    the component uses as the badge's `aria-label` and tooltip, so a
    screen-reader user was read it as though it were about them.
    
    **Two lock lessons worth keeping.** The badge already had a cell asserting
    `title === indicator.description` — true of ANY description, so it proved
    the wiring and never the words; the new cell bans renderer vocabulary
    across all five states instead. And the two spelling fixes are locked by
    ONE invariant over the member-facing copy sources rather than four string
    assertions, with comments stripped before matching (a lint-shaped test
    that fails on its own explanations teaches people to stop writing them)
    and two guards against a vacuous pass — every source must exist and exceed
    200 bytes, and a calibration cell feeds the matcher the exact strings that
    shipped and requires it to fail on them.
    
    **Two verdicts worth reusing.** A string is `not-member-facing` when the
    thing that carries it has no member-facing consumer — `DomainAction.description`
    feeds `buildAssistantTranscript`, which nothing renders; an agent tool's
    `description` is an instruction to the model. And a string is `keep` when
    it is correct copy held behind a gate that works — the greeting, help and
    unknown replies are all built from `authorizedDomains`, so their Veritas
    and Metis lines are simply absent in V1.0 and right in V1.2. **Ruling
    "unreachable" requires finding the gate and naming it**, which is how the
    persona catalog's "Metis Teacher" was measured and WITHDRAWN (the chip
    renders `familyLabel`, "Teacher"; the label needs an active Metis room)
    and how the two prompt strings were found NOT to be withdrawable.)_
    
  • 13.2 Terminology consistency table: one canonical term each for the assistant (member-facing name), tours, the audit walk, work items, decisions — grep the UI for off-terms ("Copilot" vs "assistant" vs "Lilith" on member surfaces; "Eve" must appear on NO member surface — verify with a case-insensitive sweep of rendered pages, not just source). _(CLOSED 2026-08-15 s34. Table: docs/audits/EVE_ASSISTANT_TERMINOLOGY_2026-08.md, derived from V1/BRAND.md, with a canonical term and its off-terms for all five things the task names plus room-vs-domain, and — as important — where the other word is still CORRECT (operator surfaces keep the copilot name).

    **EVE-VIS-273 (S2).** 62 member-facing strings called her "the assistant"
    against 28 that called her Lilith, usually disagreeing inside one file:
    `indicators.ts` explains persona identity twice, once as "Lilith is
    answering without one of her named guides" and once as "The assistant
    speaks generically without a named guide". **The sharpest instance was
    invisible on screen** — the shell launcher every member sees on every page
    carried `aria-label="Open AI assistant"`, silent to anyone who could see
    the button and read aloud to everyone who could not.
    
    **The task is right that source is not enough, and it cost nine more
    sites to find out.** After fixing all 62 inventoried strings, the rendered
    sweep found nine that the inventory never covered because they live in
    room and library components rather than assistant modules: "Open
    assistant" on Home and `/assistant`, "Ask assistant" ×3 in Library,
    "Memory scopes the assistant retains for you" in Profile, and "Open
    assistant with this thread" on all four room pages.
    
    **Verified on rendered pages: `eve=0` and `assistant=0` across 12 member
    routes**, measured over 22–978 visible text nodes and 22–82 accessible
    names each — and the sweep reads accessible NAMES as well as text, which
    is the only way the launcher defect was ever going to be seen. Both locks
    carry a minimum-content floor so a sweep of a gate cannot read as a pass;
    `/profile/persona` legitimately renders little without voice packs
    configured, so its floor is lower and SAID so rather than the route being
    quietly dropped.
    
    **The one exception is deliberate and encoded rather than assumed:**
    BRAND.md requires the honesty line "Responses come from Lilith, an AI
    assistant — not a human operator", so the lock bans "assistant" used as
    her NAME and keeps the common noun after hers legal. Banning the word
    outright would have deleted the one sentence that has to say it.
    
    Six tests were agreeing with the old wording and were updated with it.)_
    
  • 13.3 Microcopy for every empty/error/loading state touched in Phases 2–12 reads as Lilith (member) or operator-plain (admin) — final read-through of the inventory by surface. _(CLOSED 2026-08-15 s34. The read-through itself happened in 13.1, surface by surface, and every string carries a verdict and a note; this task is what came out of reading the empty, error and loading states as a set rather than one at a time.

    **EVE-VIS-274 (S3).** The panel's twelve in-progress labels use two
    different ellipses: eleven write `…` and one writes three periods —
    "Speaking...", on the label shown while she is talking — as did the
    truncation markers on a clipped card title and summary. Invisible
    string-by-string; obvious in a column of twelve. And the two notices a
    member gets when a turn fails were the last member sentences still
    speaking the machine's language ("the assistant isn't available on your
    current access", "trouble reaching the assistant service").
    
    **Why those two survived 13.2's rendered sweep is the reusable part:
    they only render in the refused and outage states, which a happy-path
    sweep never enters.** The rendered instrument is strictly stronger than a
    grep and still cannot see a branch it does not take — so a rendered sweep
    closes a question about the paths it walked, not about the surface.
    
    **What the lock does NOT assert is as considered as what it does.** Six
    empty states end in an exclamation ("Keep going!", "build your streaks!",
    "Start your growth journey by setting one!"), which reads as app-chirp
    beside Lilith's calm register — but that is a judgement, and a test
    encoding it would be a test of somebody's ear. Recorded here, not
    asserted. The lock takes only properties a set can be checked for: one
    ellipsis, and no machine vocabulary inside a shipped sentence.
    
    The wearable and desktop empty states ("No complication data available
    yet. Sync your watch first") were left as written: EVE-VIS-267 made them
    unreachable, and they are correct copy for the day an adapter lands.)_
    

Phase 14 — Regression hardening (lock every fix)#

  • 14.1 Every S1/S2 ledger row has a failing-then-passing regression lock (component spec, route spec, or Playwright e2e) — audit the ledger table; no row ships on a manual-verify promise alone. _(CLOSED 2026-08-15 s34. The audit is a committed checker rather than a read-through: node tools/eve-polish/ledger-lock-audit.mjs. Result: 271 rows, 192 closed S1/S2 rows audited, 0 shipping without a named lock, 24 open S1/S2 rows (no fix shipped, so no lock is owed), 2 malformed — both pre-EVE-VIS-074, which this file documents as unrecoverable.

    **Getting there needed the table repaired first, and that was the real
    find.** 64 rows — every id from EVE-VIS-001 to 077 — had been reflowed by
    prettier into a single 722-line prose block and were **invisible to any
    count of this ledger since session 11**. Session 11 saw the damage, fixed
    what it could, added `.prettierignore`, and deliberately declined to
    reconstruct the rest because "which column a given sentence belongs to is
    not always recoverable". That caution was right to state and turned out
    to be **too pessimistic in one specific way: the ` | ` delimiters had
    survived the reflow**, so rejoining the lines put every cell back where it
    belonged. 63 of the 64 landed well-formed on the first attempt; the 64th
    had one stray pipe left mid-sentence by the wrap. Done under the same
    losslessness standard session 11 used — prose identical after whitespace
    normalisation, same ids, pipe count down by exactly the one artifact.
    
    Also repaired: six rows whose regexes contained an **unescaped `|`**,
    which silently adds a column (`clear|partial|overcast|lightDome`,
    `/\b(veritas\|metis)\b/i`); one row carrying **two root-cause cells**,
    a present-tense draft and the past-tense version that superseded it; and
    EVE-VIS-075, which had lost its last four columns entirely and so sat
    outside every total — its missing cells now say what is true (the
    measurement survived inside the symptom; nobody has investigated it)
    rather than inventing an evidence trail.
    
    **One row genuinely shipped on a manual promise: EVE-VIS-007**, whose
    lock was "interaction console clean (re-swept in s5)" — a person looking
    once. It now has a real one, and **two instruments were tried and thrown
    away first, both of which would have shipped green against the restored
    bug**: React's rerender warning never fires under jsdom, and the `style`
    attribute cannot tell `border: 1px solid transparent` from the three
    long-hands because jsdom expands the shorthand into a byte-identical
    string. What separates them is `border-image`, which only the shorthand
    resets. Mutation-calibrated on the real pre-fix source.
    
    **The checker's own false positives are worth keeping in mind:** its
    first version reported the entire Withdrawn table as malformed (that
    table has four columns, not eight), and its second flagged the repaired
    EVE-VIS-007 because the new cell *mentions* the hand-check it replaced.
    A gate reporting many failures is not the same as many failures.)_
    
  • 14.2 Close the two standing e2e waivers that cover Eve: write the assistant-voice and assistant-tour Playwright suites (the curation map's named gaps); re-run the inventory miner + product graph so the new journeys enter the walk; TOTALITY spec green. _(CLOSED 2026-08-15 s34. Both suites written, both waivers removed, inventory re-mined (727 journeys / 5,927 tests), product graph rebuilt, and TOTALITY green at 32/35 flows verified + exactly the 3 remaining waived (shell.command-palette.run, nyx.reminders.set, metis.courses.enroll — none of them assistant flows).

    **EVE-VIS-275 (S3), and it was in the panel rather than the player.** A
    message sent before the panel's session exists goes to the DETERMINISTIC
    `/message` route instead of the streaming `/turns` endpoint, and
    `/message` carries no `turn.ui` intent — so no tour ever starts. From the
    outside that is indistinguishable from a broken player: panel open,
    composer working, reply arriving, tour poll timing out. Four theories were
    measured and discarded first (a drifted seeded account — real, but half
    of it; a bloated `next dev` — **0.56 GB RSS**, so no; the anchor set; a
    pre-hydration click). The failure SNAPSHOT settled it — a healthy panel
    with "Ask Lilith anything" in it and no tour anywhere is a turn that went
    somewhere else. Fixing it took the tour suite from 4–9 minutes a run to
    **~21 seconds**.
    
    **Where the doubles go is the whole design of both suites.** The tour
    suite doubles only the provider at the turns endpoint; the panel, the
    intent routing, the player, the anchor resolution and the controls are
    real. The voice suite launches Chromium with a **real fake microphone**,
    so `getUserMedia`, `MediaRecorder`, mime negotiation and blob assembly all
    genuinely run, and only the two VENDOR calls (`/audio`, `/tts`) are
    doubled — doubling `MediaRecorder` instead would have left the half most
    likely to break untested. Playback is instrumented on
    `HTMLMediaElement.prototype.play` because `new Audio(url)` is never
    appended to the document.
    
    **Two product affordances came out of it.** `TourPlayer` gained
    `data-assistant-tour-spotlight`: the tour draws its own cut-out rather
    than mounting `AnchorSpotlight`, so without it a test can ask which STEP
    the player is on but not which ELEMENT it points at — and a player
    advancing a counter over nothing looks identical. And the voice suite had
    to learn that `shouldSpeak: true` is the SERVER's request while
    `ttsEnabled` is the member's answer: its first version measured zero TTS
    calls against a panel behaving exactly as asked.
    
    All four cells calibrated: removing the `tour_start` intent reddens both
    tour cells, `shouldSpeak: false` reddens the TTS cell, and a wrong
    transcript reddens the STT cell. Three consecutive clean tour runs and
    three clean voice runs. The parity spec names `WAIVERS_CLOSED_SINCE_GOLDEN`
    rather than editing the frozen pre-graph golden, and asserts each closed
    flow is genuinely VERIFIED — a waiver deleted without a suite behind it
    would otherwise pass.)_
    
  • 14.3 Add the scroll-containment + auto-scroll + follow-up-hierarchy assertions to the existing AssistantPanel spec suite (member) and AdminAssistantChat spec (admin). _(CLOSED 2026-08-15 s34. Three cells in AssistantPanel.test.tsx, two in AdminAssistantChat.spec.tsx, all mutation-calibrated against the real pre-fix code.

    **The first thing this task turned up was that the suite it names had not
    run in months — EVE-VIS-276 (S1).** `AssistantPanel` calls `useAuth()`,
    which throws outside an `AuthProvider`, and none of the three specs that
    render the panel provided one: **all 77 tests failed at render**, about a
    millisecond each. Twenty-seven ledger rows name `AssistantPanel.test.tsx`
    as their lock and not one was running. Two sessions had been recording
    those 77 as "pre-existing failures verified against the stashed tree" —
    true every time, and exactly how a dead suite stays dead. **A stash check
    answers "did I break this"; it never answers "is this supposed to be
    red".** One mock restored all 77.
    
    **jsdom computes no layout, so none of the five cells measures pixels** —
    each asserts the MECHANISM that regressed. Containment is stated as the
    API the fix must not use: `scrollIntoView` scrolls every scrollable
    ANCESTOR, which is what yanked the page behind the panel, so the cells
    INSTALL `scrollIntoView` and `window.scrollTo` (jsdom has neither, so
    `vi.spyOn` throws) and require them untouched. Auto-scroll observes the
    `scrollTop` assignment the implementation falls back to when
    `Element.scrollTo` is missing. Hierarchy asserts follow-ups render INSIDE
    the assistant's message and after its text.
    
    **Three instrument faults, each caught by calibration or by the suite.**
    (1) The containment cell PASSED against a panel reverted to
    `sentinel.scrollIntoView(…)`, because it asserted before the scroll
    effect had run — "not called yet" and "never called" look identical, so a
    negative assertion needs the positive one in front of it. (2) The
    follow-up cell read the panel's own opening GREETING chips, which
    `querySelector` finds first; it takes the last. (3) The new cells broke a
    LATER describe while passing alone — the panel persists its session id,
    so the next describe's panel resumed `sess-14-3` and rendered nothing.
    Passing alone and failing in the suite is the signature of leaked state,
    and the state is rarely the mock.
    
    The admin side deliberately does NOT get a follow-up cell: the cockpit has
    no suggestion chips, and inventing one to make the two suites symmetrical
    would be a test of nothing. It gets the ordering invariant instead.)_
    
  • 14.4 Visual regression pilot: adopt Playwright screenshot assertions for the 5 most defect-dense surfaces found in this initiative (likely: panel conversation state, tour dialog, audit board, explorer explore-lens, admin drawer) — masked dynamic regions, both themes. _(CLOSED 2026-08-15 s34. assistant-visual-regression.spec.ts, 10 shots = 5 surfaces x 2 themes, 10/10 twice at ~1.6 minutes, baselines committed beside the estate's existing -chromium-darwin ones.

    **The five are MEASURED, and the measurement corrects the task's guess.**
    Counting the ledger by surface: member panel 105, room surfaces 30, tour
    20, shell chrome 19, audit board 12, admin drawer 11, graph explorer 5. So
    room surfaces and shell chrome are in; the explorer is out — it is also
    builder-gated, so a member-session screenshot of it would be a picture of
    a permission gate. The admin drawer lives in another app with its own
    runner and is left for a sibling pilot rather than smuggled in.
    
    **`maxDiffPixelRatio` is a fraction of the shot's AREA, so one threshold
    cannot serve every shot — and finding that out was the whole exercise.**
    Reverting the tour card's 14px corner radius to 0 does NOT fail it at the
    estate's usual 0.02; lengthening the composer placeholder does NOT fail
    the panel at 0.02 either. Measured and set per shot: tour card 0.001
    (catches the radius), panel and chrome 0.004 (catches the placeholder,
    ~3k pixels against ~320k), and the two FULL-VIEWPORT shots stay looser
    because ~3k against ~1.3M is 0.002 and nothing sane would catch it. The
    header says so out loud — those two catch layout shifts, colour changes
    and things disappearing, not a reworded label. **A gate that cannot fail
    is the default outcome here**, and 0.02 on a clipped element is exactly
    that.
    
    **Two stability traps, both already known and both re-met.** The theme is
    VERIFIED, not assumed — `cream` is the absence of `data-lilith-theme`, so
    a pass that never applied the attribute is indistinguishable from a
    correct cream pass. And the tour cell alone takes a FRESH seed: the other
    four want a stable account because its content is part of the picture,
    but a stable account carries a session the panel RESUMES instead of
    running the scripted turn — so that cell kept photographing a panel with
    no tour in it. Transitions are zeroed before every shot; the greeting
    (time-of-day dependent), the diagnostics block and the account chip are
    masked, because their changing is not a regression.)_
    

Phase 15 — Final adversarial polish gate (the definition of done)#

  • 15.1 Fresh session, fresh member account, full re-run of the HAPPY-PATH SPINE end to end in one sitting: signup → onboarding → panel conversation in 2 domains → a do-tier confirm → a full curated tour → voice turn → audit begin + 3 flows + board → (admin) log issue → approve card → read tools → explorer inspection of the new item. ZERO new ledger entries allowed. Any finding reopens the relevant phase. (2026-08-15 s36: run 15 of phase-15-1-happy-path-spine.spec.ts — member eve-spine-632195 created by the run, all member stages, SPINE FINDINGS 0; phase-15-1b-operator-loop.spec.ts on the admin app — log-issue → approve verified in Postgres work_item rows → read tools, OPERATOR FINDINGS 0; phase-15-1c-explorer-item.spec.ts back on member web — the board holds the item, the work lens paints its title, the card hit-tests to itself, EXPLORER FINDINGS 0. One sitting carried by one BFF process + one Postgres + the run-id state files. Runs 12–14 were the diagnosis runs: their findings produced EVE-VIS-281/282 (both fixed and closed BEFORE the final walk), the announce-before-act wire hold, the refusal log line, and three instrument repairs — store-paced polls under the 60/min rate limiter, settle-before-send, stage-4 retries. The curated-tour residual under the harness model stays ledgered at EVE-VIS-280 with per-run telemetry.)
  • 15.2 Adversarial screenshot review: a second pass over every Phase-15.1 screenshot deliberately hunting what the first pass normalized (the "fresh eyes" rule: inspect in reverse order, zoomed). CLAUDE.md two-pass discipline applies to pixels too. (2026-08-15 s36: all 22 evidence screenshots read in reverse order (21 → 01), each examined as an image. ZERO new product defect rows. What the hunt caught, classified: (1) a bottom-left "N" circle overlapping the admin sidebar's last nav rows in all four operator shots — Next's dev-tools badge, dev-only chrome; INSTRUMENT, and 15.1b now neutralises it like the member specs do; (2) 14-audit-begun.png captured the panel mid entrance-fade (ghost bubbles) — capture timing, the same transcript fully painted in 15/16; (3) model-prose blemishes under the harness model, labelled per the cost rule: a "tâble" typo in the admin copilot's reply, and "favourites" in Lilith's prose mirroring the member beside the product's own "favorites" strings (product copy consistent, 098's standard holds); (4) "Generated 3/23/2026" on the admin Copilot-health card — the dev seed's own stamp, data-true rendering. Positive finds the reverse order surfaced: the wire hold visibly working (16: state pulled and accurate BEFORE "Marking it now"), the confirm-first discipline in the model's own words (08), and the CURATED shell-orientation tour running 3/3 steps in run 15 (12).)
  • 15.3 Console-clean certification: every route touched in 15.1 shows zero errors/warnings in both themes. (2026-08-15 s36: phase-15-3-console-clean.spec.ts, run against a PRODUCTION build (next build --webpack + next start) so dev chrome could not be what got certified — 10/10 member cells green (cream and dusk × /welcome, /, /domains/tara, /assistant/audit, /assistant/graph?lens=work with the operator credential 15.1c used) plus the admin operations dashboard, all with ZERO console errors/warnings and zero uncaught exceptions. Getting the production build to EXIST was the bulk of the task and closed two found-in-passing rows: EVE-VIS-213 (onnxruntime-web undeclared; then its second act — the config's own alias assumed the hoisted linker; then webpack pinned because Turbopack refuses workbench-kit's NodeNext specifiers) and two fresh client-graph breaks fixed the same night (the V2 shell tile importing the persistence BARREL dragged pg into the client bundle; the Veritas story workspace's barrel import dragged node:fs through two dead sub-barrel re-exports, both removed with zero external consumers). Font-preload hints (EVE-VIS-043's condition): ZERO on the production build across all ten cells — the warnings were dev-only next/font behaviour, and that row closes with this measurement.)
  • 15.4 Ledger closure audit: read every row; every fix commit exists, every regression lock runs in CI, no row in "open". The ledger's summary header states totals by severity with zero open. (2026-08-15 s36: zero open rows across ALL severities, verified with the checker's own cell parsing — the last two open rows closed tonight (EVE-VIS-043 on its own production-build condition, measured at zero hints; EVE-VIS-213 with the build actually building). Lock audit: node tools/eve-polish/ledger-lock-audit.mjs → "OK — every closed S1/S2 row names something that runs" (219 closed S1/S2 audited). Fix-commit sweep: every 10-hex commit hash cited in any row exists in this repo (23 of 23; two lookalike hex strings are string-catalog ids inside evidence prose, verified by reading them in place). Lock-file sweep: every spec path a lock names resolves on disk (one stale citation found and fixed — 089's lock had been renamed .test.ts→.spec.ts by the conventions ratchet the same evening). The summary header now opens the ledger with the totals: 279 rows — S1 68 · S2 133 · S3 49 · 2 no-severity found-in-passing · 23 legacy-shaped (closed) · 4 withdrawn — zero open.)
  • 15.5 Update the design doc (§ new: "Polish gate 2026-08") and memory with the outcome; commit + two-line push; mark this file's acceptance block below. (2026-08-15 s36: V1/OMNIPRESENT_ASSISTANT_DESIGN_2026-08-03.md gained § "Polish gate 2026-08" — the gate's outcome, the completion-checker family the initiative left in the product, and the residuals stated honestly; auto-memory updated (eve-deep-polish-initiative → COMPLETE, with the false-defect traps and prod-build landmines for the next session). This checkbox travels in the closing commit, pushed to branch and main — the two-line push is the last act of the initiative.) (Completion re-audit 2026-08-30: that named external auto-memory is not retained in the current environment, so its historical content cannot be authenticated. The versioned sibling handoff for the subsequent capability-breadth initiative is now docs/agents/eve-everywhere-initiative.md; this pointer is the durable replacement, not a claim to have recovered the missing note.)

Acceptance sign-off (2026-08-15 s36)#

All seven criteria verified this session: (1) every checkbox above is [x], each carrying its own session-stamped verification note; (2) the ledger holds zero open rows across all severities, checker green, every close naming a runnable lock; (3) the visual phases each recorded both-themes × three-viewports evidence in their own verification notes (the matrices are named per phase); (4) spine run 15 produced zero new findings, and so did the operator and explorer halves; (5) assistant-voice + assistant-tour suites exist and TOTALITY is green at 32/35 verified with exactly the three non-assistant waivers (closed at 14.2, corroborated live in the 15.1c evidence); (6) the member-surface vocabulary sweeps (13.2/13.3) are closed with their locks; (7) this commit and its push to branch AND main carry all of it.

Acceptance criteria (ALL required)#

  1. Every checkbox above is [x] with same-session verification.
  2. Defect ledger: zero open rows, all severities, every row regression-locked.
  3. Both themes × three viewports evidence exists for every visual phase.
  4. The Phase 15.1 spine run produced zero new findings.
  5. assistant-voice and assistant-tour e2e suites exist and are green in the composed inventory (waivers removed, TOTALITY green).
  6. No member surface renders internal vocabulary (Eve, handoff/mode-state debug chrome, provider/error codes).
  7. All work committed and pushed to branch AND main.

Non-goals (so scope cannot silently creep)#

  • Non-assistant product surfaces beyond the Phase 12 list (the July UI/UX audit owns the wider estate).
  • Model-quality tuning beyond flagging weird replies (prompt/provider work is its own initiative; ledger model-attributed findings for it).
  • Realtime speech-to-speech, avatars, MCP federation (design-doc "later" items).