# Screenshot verification follow-up

This repairs the specific web typecheck failure recorded by the
[execution review](../agentic-execution-source-review-2026-09-07/README.md). It
does not complete the presentation library or the full Eve release gate.

## Change and design boundary

The page-screenshot helper already rejected hidden pages before capture and
after rendering. TypeScript retained the initial `visible` narrowing across the
awaits and rejected the second comparison. A small boolean helper now performs
each fresh read, preserving those runtime checks without a cast. The existing
opt-in control, panel styling, preview, fallback copy, redaction, size limits,
model contract and narration inputs are unchanged.

The frontend skill informed this deliberately limited visual scope: retain the
existing panel and preview hierarchy; keep consent, capture status and result or
honest fallback as the content sequence; preserve existing transitions and test
interruption instead of adding motion or chrome. The desktop and mobile
synthetic captures were visually inspected: public chart bars remain visible,
while the marked private regions, input values and chat are absent. DOM
rendering still differs from native browser pixels; this is not a new visual
fidelity claim.

## Reproducible setup

`html2canvas-pro` was already pinned to **2.4.1** in the web manifest and lock.
An offline filtered install failed because that exact tarball was missing. The
following succeeded without changing any dependency manifest or lockfile:

```sh
pnpm --filter @oshun/web install --frozen-lockfile --ignore-scripts --prefer-offline
```

The first browser attempt passed the chart case, then Next dev restarted during
route warmup. Its installed implementation restarts above 80% of the heap limit;
the focused configuration's 4 GiB heap was insufficient. The verified owned
runner was interrupted and cleaned up, and the initial log is retained with its
ephemeral test-session cookie headers removed. The focused gate now uses a
bounded **5 GiB** web heap, one worker and no retries. Its complete 16-case
rerun passed without a restart. The resource supervisor's lowest sample was
**3.3 GiB available**, above its 2 GiB termination reserve. The owned narrator
was paused during browser execution and resumed after both attempts. Other
checkouts and their processes were not altered.

The focused test TypeScript configuration lives in the web app so its ambient
types resolve correctly. It binds `@playwright/test` to the web consumer's
locked **1.60.0**: the shared accessibility package has a type-only import and
separately resolves **1.57.0** in the lockfile, but receives the web runner's
test object at runtime. This preserves typed callbacks and matches the actual
browser run; it neither changes dependency versions nor disables compiler
checks. The two other aliases retain the web source and configuration
entrypoints. Remaining workspace packages resolve through their declared
exports.

An earlier temporary configuration outside the app could not resolve its ambient
types. The baseline ratchet incorrectly reported zero app-owned errors because
the compiler's pathless TS2688 diagnostics do not match its file/line filter.
The raw compiler exposed that false green. That interim result is **not accepted
evidence**. The app-local focused configuration is checked directly with `tsc`;
hardening the shared ratchet remains a separate, explicit checklist item.

## Verification scope

- **51 web tests** in five suites: screenshot admission and late visibility,
  navigation and abort results; PNG failure cleanup; redaction; consent reset;
  close/session/unmount cancellation; concurrent capture; text-page fallback;
  streaming and ephemeral reply retention. Fourteen helper/hook cases are new.
- **141 BFF tests** in seven suites: page vision and screenshot contract,
  client-tool bridge, tool availability, reply inspection, trust labels, threat
  model and assistant turns. These include cross-member/session rejection and
  keeping image bytes out of the main model request and transcript.
- **97 V1 prompt-boundary/corpus tests** in two suites. Providers are controlled
  test implementations, not live model accuracy or resistance measurements.
- **16 desktop/mobile Playwright tests**, including shared serious/critical WCAG
  2.1 AA checks. Four new project cases interrupt the real renderer after its
  iframe appears: synthetic visibility change or the actual disable button. They
  assert error-only delivery, no image preview, iframe cleanup and preserved
  live form values. Existing cases cover actual public/private pixels, byte and
  dimension limits, permission reset, reply details, keyboard focus, capture
  recovery and provider-failure provenance. Auth, session ownership, SSE and
  client-tool exchange are real; the chat/vision providers are test doubles.
- Serial affected typechecks for **web, BFF and prompt-injection-evals** passed.
  The existing web/BFF ratchets report zero app-owned errors; they do not
  certify every transitive dependency. No baseline was widened.

Run the strict focused test compiler and the complete browser matrix from root:

```sh
pnpm --dir apps/oshun/web exec tsc --noEmit -p tsconfig.page-inspection-tests.json
pnpm exec playwright test --config apps/oshun/web/playwright.page-inspection.config.ts
```

Check resources first and keep expensive verification serial. The browser gate
starts the isolated test BFF and Next dev servers on 4010/3010; it needs no
database migrations or paid provider call. Compilation is verified through Next
dev. No Next production build or broad workspace sweep was run under this host's
memory-safety policy. Mobile here means browser viewport emulation, not native
mobile application automation.

`state.json` binds named source/configuration inputs and the retained logs. Run
`python3` on the sibling `verify.py` to check those hashes, and on
`test_verify.py` for positive/missing-input/changed-input/changed-evidence
regressions. This is a bounded entrypoint receipt, not an exhaustive transitive
dependency manifest. The earlier execution receipt intentionally retains the old
helper hash and is historical after this repair; its failure evidence was not
rewritten to appear current. The unrelated root discovery/dependency/baseline
findings remain open, as do presentation source acceptance, global audio
verification and reader review.
