Implemented all three modes, offline Ghost Race, local progression, original art, and the IPFS-ready dist/ export.

Build, typecheck, 13 tests, Chromium/WebKit flows, and 16/16 design checks passed. Submission: 1.25 MiB / 8 MiB. Live transitions used fixtures while the real wheel rested.

README · Validation and screenshots

Not delivered: artifacts/accessibility.json, artifacts/browser-results.json, artifacts/design-guidance-license.txt, artifacts/design-probe.json, artifacts/export-manifest.json, artifacts/screenshots/home-1440.png, artifacts/screenshots/home-390.png, artifacts/screenshots/predict-390.png and 4 more. Under artifacts/, only this step's named outputs leave the machine.

Saved the review to .imd-findings.json. Website files remain unchanged.

• High — bug: real measurement milestones break the events parser. One valid event without confirmationStatus discards the entire feed. Milestones disappear, and predictions remain pending despite available settlement evidence. Parse measurement and chain events separately, preserving confirmation checks for settlement scoring.
• Medium — bug: connection notices shift primary controls by 76 px. At 390 × 844, successful loading moves both the race arena and prediction form upward. Keep the status row mounted and update its contents without moving the controls.

Build, 13 rule tests, 12 automated accessibility scans, layout checks, and Chromium/WebKit race flows passed. Three missing clauses below share the events-parser defect.

Both reported defects were real, and both are fixed. dist/ has been rebuilt from the final source, and my answer to each finding is in .imd-responses.json.

1. The events feed failed on real milestones (high). This happened exactly as reported. The live feed's daily distance milestone has no confirmationStatus, so the parser threw and all twenty events were thrown away.

• Parser (src/api.ts): events are now read by source.kind. Chain events still must have a confirmation status; measurement events don't. A single bad event is now skipped on its own instead of sinking the whole feed.
• Prediction scoring (src/predictor.ts): only safe or finalized chain settlement events count as evidence, so measurement events can never resolve a prediction.
• How to play (src/HowToPlay.tsx): the milestone row shows whichever supported event is newest. Your captured replay now shows "1,000 m in one day · Daily distance milestone · Oct 31 UTC"; with the live API it showed "Round #769 settled · 495 turns added".
• Tests: I added two using a trimmed copy of the real response (tests/fixtures/events-measurement.json). One checks the milestone parses while safe settlements stay usable. The other checks that a pending round-766 prediction of 0 turns now resolves from the events feed alone.

2. The race and prediction controls jumped 76px (medium). This also happened. The connection notice is now always shown and only its text changes: "Checking the wheel", "Wheel connected" (with a "Check now" button) or the offline message (with "Retry feed"). Its height is fixed at 60px on phones and 64px on desktop. I replayed your scenario in Chromium: holding the API, releasing it to the live API, then cutting the connection. The race area stayed at y=254px (390 wide) and 334px (1440 wide), and the prediction form stayed at 359px and 334px through all three states.

Checks run:

• Typecheck: passed.
• Tests: 15 of 15 passed.
• Production build: passed.
• The project's full browser check: all passed. This covers all six pages at four widths, accessibility scans, offline play, the live race and the predictor.
• Export check: passed, at 472 KB.
• Design probe: 16 of 16 passed, and the 390-wide first screen is unchanged.

Limitations and other changes:

• WebKit isn't installed on this machine, so every browser check ran in Chromium only.
• The README used to link to artifacts/validation.md, but that file wasn't in the tree. This machine's git settings exclude the artifacts/ folder, so it may never have been committed. I put the actual command results in the README's validation section and also wrote artifacts/validation.md.
• I updated the FeedNotice entry in DESIGN.md to match the new behaviour.
• Nothing is committed.

Not delivered: artifacts/export-manifest.json, artifacts/validation.md. Under artifacts/, only this step's named outputs leave the machine.

Both previous findings are fixed at their original reproduction widths, confirmed in Chromium and WebKit:

• Measurement milestones render correctly. Round 766 resolves from safe event evidence alone to 0 turns and 525 points.
• The connection notice remains mounted. At 390px, the race arena stays at y=254px and prediction form at y=359px through loading, success, failure and recovery.

One residual defect is recorded in .imd-findings.json.

Medium — bug: round-schedule text still shifts prediction controls on narrow phones

At 360×844, hold the initial API requests on #/predict, wait for fonts, then release successful responses. The schedule banner shrinks 85→67px, moving the number input y=518→500px and submit button y=757→739px. At 320px, they move 42px. Connection loss reverses the movement.

Reserve the banner’s measured maximum height: 85px at 360–389px; 109px below 360px, preserving current 390px/desktop sizing. Browser-only application of these values kept the controls stationary through all four states. Source: styles.css:216.

Strict TypeScript, production build and all 15 rule tests pass. Rebuilt files match dist/. Chromium checks passed across all routes, including 12 automated accessibility scans; WebKit passed offline racing, keyboard/touch input, pause/resume, finishing and profile persistence.

The requirement ledger below counts related clauses together where they share one implementation.