Add camera-first, offline-capable GridPuzzle phone app

[senegrom]

Camera-first GridPuzzle phone app

browser-scanner hosts the installable PWA deployed at https://senegrom.github.io/GridPuzzle/. Nothing in this PR merges automatically into master.

What it includes

• Rear-camera capture, stable-grid auto capture, photo import, perspective correction and grid-size/type suggestions.
• Local printed-clue OCR; photographs are not uploaded to a remote recognizer or solver.
• All twelve GridPuzzle families through a strict row-major JSON adapter and the complete Python 3.14 solver in a cancellable Pyodide worker: Sudoku, Killer Sudoku, Futoshiki, KenKen, Latin square, diagonal/pandiagonal Latin square, Hidato, Numbrix, Kakuro, Slitherlink and Str8ts.
• Exact native Str8ts street constraints. Solid black cells split streets; numbered black cells still participate in row/column uniqueness without joining a street.
• Editors for values/blocked or black cells, cages, inequalities and Kakuro directional clues; undo; guided Review highlighted clues → Save & next; JSON import/export.
• Unique/multiple/no-solution/invalid/unfinished states, time limits and real cancellation.
• Clean-board and captured-photo solution overlays, including Slitherlink edges.
• Installable icons, saved puzzle/settings, hash-verified offline preparation and self-hosted Pyodide/Tesseract assets.

Recognition trust and real newspaper regressions

Automatic recognition is a proposal, not proof. Automatically identified rules require confirmation, structural families remain review-gated, and wrong/missed OCR is never silently converted into solver certainty.

Newsprint processing now distinguishes true solid black cells from gray Sudoku shading, suppresses paper/halftone specks with connected-component filtering, and locally Otsu-binarizes digit crops before the single bounded Tesseract atlas call.

The two user-provided newspaper examples from 2026-09-07 are committed under Examples/BrowserScanner/Newspaper/, outside web/ so they are not shipped in the PWA/offline bundle. The Pages gate runs them through the production scanner in Chromium and WebKit.

Current raw results on the tested deployment artifact:

fixture	Chromium 153	WebKit 26.6	unsafe discrepancies
shaded Sudoku	24/24	24/24	0 / 0
Str8ts	19/20	19/20	0 / 0

For Str8ts, both engines detect the exact 22-cell black layout. The single missed printed value is explicitly flagged for review in both engines. Generated regressions are also green: Chromium reads every tested generated variant exactly; WebKit reads the perspective/shadow case 29/30 with its miss flagged.

Phone/offline hardening

• Recognition uncertainty persists atomically; photos do not.
• Malformed imports and structural inconsistencies are rejected before rendering/solving.
• Cage/Kakuro solve-ready validation runs before Pyodide loads.
• Keyboard boundaries and no-op edits use authoritative puzzle state.
• Large phone-photo decoding is bounded/downscaled.
• CSP/no-referrer and iPhone safe-area handling are in the core app.
• Offline runtime assets are content-addressed and digest-verified; startup readiness is cheap while explicit offline preparation performs full verification and requests persistent storage.
• Query-string root navigation works offline.

CI and deployment

The existing Build and deploy phone scanner workflow is the single expensive Chromium/WebKit deployment gate. It runs the complete native bounded suite, browser unit tests, all twelve solver families through real Pyodide, generated OCR/review tests, the two real newspaper OCR regressions, uploads the tested static artifact, and publishes that exact artifact to Pages. Lightweight PR browser CI avoids duplicating the full gate; Linux/Windows CI and forward compatibility remain separate.

No temporary Str8ts apply/retrigger workflows or staging helpers are present on browser-scanner; the temporary .str8ts-stage payloads were removed before the clean integration commit.

Remaining empirical limit

The strongest remaining validation gap is physical-device testing: actual iPhone autofocus/exposure, installed-PWA camera behaviour, storage eviction and true airplane-mode use. Generated and desktop-engine tests are not presented as substitutes for that hardware testing.

See web/README.md, web/TESTING.md, Examples/BrowserScanner/Newspaper/ground-truth.json, and the uploaded browser test artifacts for the current methodology and measurements.

[senegrom]

Final verification — 050852f74804

Implementation is pushed to browser-scanner; nothing has been merged into master. Relative to the starting commit, the branch adds 26 files and does not modify any existing solver file, technique hierarchy or search profile.

The build job passed in run 34052179905:

• 626 native Python tests passed, plus 22 JavaScript tests.
• 23 recorded acceptance checks passed in each browser: Chromium 153.0.8010.12 and mobile WebKit 26.6. These exercise the real Python/OCR WebAssembly runtimes, all eleven solver families, editing/undo, type overrides, cancellation/restart, persistent uncertainty, camera-denial fallback, photo overlays and offline operation.
• No uncaught browser errors or external runtime requests were recorded.
• Offline checks physically stopped the origin server, verified it was unreachable, then reloaded the cached app, started a fresh Python solver and performed fresh photo recognition. Physical iPhone camera/airplane-mode/installation testing remains outstanding.

Recognition is still a review-assisted prototype

On the generated 30-given printed Sudoku fixture, raw recognition correctly read 25/30 clues in Chromium and 24/30 in WebKit. The five/six discrepancies were flagged and corrected through the real cell editor before checking the exact original solution. These correction counts are recorded separately in results.json; they are not presented as automatic recognition success. Cage boundaries, inequalities, path variants and Kakuro clue directions remain experimental and require review.

Deployment blocker: one owner setting

The tested static site and Pages artifact have been uploaded. The workflow's overall failure is solely the configure job: Pages is not enabled. Initial enablement returned HTTP 403 Resource not accessible by integration.

Set Settings → Pages → Source: GitHub Actions, then re-run the failed jobs in the run linked above. If the github-pages environment has branch restrictions, allow browser-scanner. No merge to master is necessary. The intended address is https://senegrom.github.io/GridPuzzle/, but it is not live yet.

Artifacts: scanner-static-build, browser-test-report, and the tested Pages upload. Implementation/limitations are documented in web/README.md and test methodology in web/TESTING.md.

[senegrom]

Starting a complete follow-up pass on the actionable review findings. I’m keeping this on browser-scanner and will preserve solver solution space/technique semantics. Main targets: automatic-scan trust, query-string offline navigation, keyboard movement, no-op edits, earlier structural validation, iPhone safe areas/contrast, offline update continuity/per-request hashing, duplicate browser CI, and stale deployment docs.