`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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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 - [x] 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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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'`.)_ - [x] 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.)_ - [x] 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 `

`, 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.)\_ - [x] 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 - [x] 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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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".)_ - [x] 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 - [x] 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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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.)_ - [x] 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 - [x] 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.)_ - [x] 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.)_ - [x] 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 `

`. 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 `