Completed the PLATFORM-UX-V1 integration handoff.

• Added integration/acceptance plan covering all three surfaces, six functions, shared component contracts, state matrix, prior three A1 findings, harnesses, WebKit diagnostic, and fresh acceptance gates.
• Added baseline hashes for commit 878ba4…1207 and all 26 protected A1 files.
• Added stage check results.

Checks: protected hashes/path diff passed; SDK tests passed (39); Forge formatting and ABI consistency passed. Forge build did not complete within the command wrapper’s compilation cutoff, and frontend/browser checks were not run because protected frontend/node_modules/ is absent. No protocol, dependency, config, or frontend source was changed.

Implemented the English platform, eight shared component families, sequence-first workspace, signing cancellation, recovery and file fixes. Source and dist/ are delivered.

Passed: build/typecheck, 41 unit, 50 browser, 125 contract and 39 SDK tests. A1 remains unchanged. Bundle: 8,336,864 bytes, under 8 MiB.

Validation and evidence retain the WebKit RPC qualification. Live compatibility remains NOT_RUN; production readiness is false.

Completed the English platform integration and static export.

Local checks passed: 43 frontend units, 175 browser cases, and 64 actual local-contract scenarios. Protected core/dependencies remain unchanged. Bundle: 8,352,738 bytes, below 8 MiB.

Evidence and limitations. Live compatibility remains NOT_RUN; productionReady:false.

Review complete. The findings file at .imd-findings.json holds three findings, ranked by severity. No tracked file was changed.

What I verified and found sound

• Protected A1 paths (src, sdk, cli, abi, config, lib, root build files) are byte-identical to baseline 878ba4b.
• Final evidence hash manifests match the tree: 124 source entries and 8 dist entries, zero mismatches or missing files.
• Browser matrix report shows 175 expected, 0 skipped, 0 unexpected. All four local signing runs exit 0, including WebKit desktop and WebKit mobile, so the prior WebKit gap is closed in the final logs. Fixture, actual-local and public results stay separately labelled, with liveCompatibility NOT_RUN.
• Fixtures reach real routes nowhere: the fixture adapter is constructed only under the demo hash route, and the local adapter refuses non-loopback pages and non-31337 chains.
• Anchor, proposal, parent job, spender, price, body hash and exact allowance are re-read inside the prepare callback before every signature, and the prepare is run twice with calldata equality enforced. Close, unmount and account or network change invalidate the action scope before the wallet is called.
• Status-1 retry requires identical body hash and phase 4. Unknown outcomes keep the original hash and nonce and block repurchase. The HTML preview is an opaque sandbox with a network-denying CSP, Markdown is sanitized with ids and names stripped, downloads use the descriptor basename and exact bytes, and PNG natural aspect is preserved in CSS.
• Token contrast: action on white 5.06, secondary on canvas 5.12, action on selected 4.58. All pass AA for text.

Findings

1. Medium, Workspace.tsx:62. A journal row whose reconciliation is final (reverted, or replaced by an unrelated transaction) is rendered as a pseudo-attempt numbered activeAttempt+1 for the same proposal. That makes the real refused attempt look superseded, so the "Review identical paid retry" button never appears, while other funding is rejected by phase 4. The contract and adapter both permit the retry. Only clearing browser storage by hand restores it. The existing replacement scenario stops at asserting the REPLACED_OTHER text, so no test covers this.
2. Low, chainAdapter.ts:32. Persisting a later non-purchase transaction that reuses a dropped purchase's nonce deletes the original purchase hash from local storage, so after reload that hash is no longer listed or reconciled. Chain truth prevents a rebuy, so this is a record-retention defect only.
3. Low, PROTOCOL-OVERVIEW.md:206. A Chinese button label remains in a current operator doc and does not match the shipped English control.

Recommendation

Reopen for the medium finding with a code fix (exclude finally-reconciled journal rows from attempt numbering, or scope isLatestAttempt to on-chain attempts) and a rerun of the local signing suite on both engines with a scenario that retries after a reverted or replaced purchase. The two low findings can ride along. Nothing here touches the protocol, and no liveness claim changes.

Saved four substantiated findings in .imd-findings.json. Implementation files remain unchanged.

• High — bug: long histories overlap sequence labels. After 18 accepted rounds, buttons measure 62.20px at 1440px and 36.95px at 768px; text spills into neighboring rounds. Preserve readable item widths in a bounded, visibly scrollable rail.
• Medium — accessibility: proposal errors lack field association and focus. Submitting empty fields leaves focus on “Review proposal”; neither invalid field exposes its error. Associate errors with their controls and focus the first invalid field.
• Medium — bug: pending reads claim empty histories. Holding a real proposalCount() response displays both “Reading proposals…” and “No proposals yet.” Show empty messages only after successful empty reads.
• Medium — bug: mixed-read acceptance depends on RPC timing. One unchanged-source run failed with two anchor reads where the test expected three; its repeat passed. Inject one transient inconsistent pair independently of unrelated proposal reads.

Fresh checks passed frontend typecheck/build, 43 units, all 175 configured browser cases across runs, and all 16 actual local signing scenarios in Chromium desktop and WebKit mobile. WebKit mobile’s fixture cases completed as 34 plus one isolated case. The initial Chromium signing failure remains recorded above; its repeat passed. All 42 protected baseline files, 120 checked source entries, and eight export files match their manifests. Live acceptance remains NOT_RUN.

“Missing” below includes partially satisfied requirements affected by the findings.

Review complete. The findings file holds three entries and no repository file was changed.

Prior findings, settled

• Hidden retry after reverted/replaced journal row (medium): fixed. The new isLatestAttempt in frontend/src/ui/recovery.ts excludes pseudo-rows whose purchase stage is "failed", and chainAdapter marks REVERTED and REPLACED_OTHER rows exactly that way. I executed the function directly under Node: a failed pseudo-row no longer supersedes attempt 1, while a DROPPED_OR_UNKNOWN row still blocks, which is the intended guard. Kept only as a low note because no test exercises the repaired path.
• Nonce reuse deleting the original purchase hash (low): fixed. Persistence now goes through mergeJournal, which replaces rows by hash instead of account/chain/nonce. Run standalone, a same-nonce propose row coexists with the original sponsor row, and a same-hash state update still replaces in place. Reconciliation then reports the original as REPLACED_OTHER because no on-chain attempt matches it.
• Chinese label in PROTOCOL-OVERVIEW.md (low): fixed. The doc now names the shipped control, and a CJK grep over docs and frontend source returns nothing.

Verified clean

• All 26 protected A1 hashes in BASELINE.json match the tree, and the protected directories have no extra files and no diff against the baseline commit.
• The fixture adapter is mounted only on the demo route in App.tsx; real routes construct the chain adapter.
• Revision-1 Chromium logs show 105 e2e tests and 16 local-contract cases passing on the final tree.

New findings, ranked

1. Medium, evidence: dist-sha256.json and export-rebuild.json describe an export that no longer exists. The final commit rebuilt dist with new asset names, and index.html now hashes differently from the recorded value. The README cites that record as the rebuild PASS.
2. Medium, evidence: WebKit desktop and mobile, Chromium-mobile local-contract, and unit logs were last produced at the previous commit, before the final six-file frontend change. The revision-1 rerun is Chromium-only and is not referenced by any evidence doc, so the prior WebKit gap remains unexercised against the shipped code.
3. Low, test coverage: the repair for the hidden retry has no regression test. The only replacement scenario removes the journal row at browser-chain.mjs line 474 before funding again.

Neither medium item is a code defect, and nothing here is high or core. Acceptance should wait on a rebuilt hash record and WebKit/mobile/unit reruns indexed from the final tree.

Updated the root README and current docs/ guides for the English Open Workshop product, including setup, routes and permissions, design tokens and components, operation and contributor guidance, and future brief language policy. The rejected frontend/DESIGN.md path was not modified.

The documentation path/link check and git diff --check passed. The pinned A1/core/build/dependency hashes also match. The read-only runtime audit fails: seven frontend source/test files and four recorded export paths differ from the retained final manifests. I recorded the current hashes and marked current exact-source frontend, browser, local-chain, and rendered-design acceptance as NOT_RUN; earlier results remain historical. Production readiness remains false and live compatibility unverified.

All four findings from the last round are fixed. I found one new issue, a low-severity taste finding, and it is the only entry in .imd-findings.json. I changed no repository files.

Prior findings, settled

• [high] Long histories collapse the sequence labels: fixed. The sequence row now gives each round at least 196px and scrolls sideways, with "Scroll sideways to see all 19 checkpoints" and Earlier/Later buttons (frontend/src/styles.css:90-93). I ran 15 funded rounds in Chromium (19 checkpoints, no page errors):
  • At 1440 and 768, every round is at least 195×104px.
  • At 390 and 320, the rounds sit in two columns and there is no horizontal scroll.
  • Every checkpoint can still be selected, in numeric order.
• [medium] Proposal errors not tied to their fields: fixed. I clicked "Review proposal" with the form empty:
  • Focus moves to New asset name, which reports aria-invalid=true and aria-describedby=target-err, with the error text beside it.
  • Requirement reports aria-invalid=true and aria-describedby="requirement-err req-len".
• [medium] Unfinished reads shown as empty histories: fixed by reading the code. Both empty messages now appear only after their read succeeds with no results (proposals.isSuccess && proposals.data.length === 0 and the same for funding attempts, Workspace.tsx:79,87). While loading or after a failure, only the loading or error-and-Retry line shows. I did not re-run the delayed-RPC reproduction against local contracts.
• [medium] Mixed-read test depends on RPC timing: fixed by reading the code and the recorded run. The test now corrupts only the anchor call that belongs to the branch read (it matches the goal call's id + 4). It expects exactly one bad pair and then checks the separate always-wrong Retry case (browser-chain.mjs:587-610). The recorded Chromium run (revision-1/local/chromium/browser-chain.json) shows the new counters (corruptedPairs 1 then 3) and result: PASS. I did not re-run the suite myself. The final/ evidence files still record the old counters, without corruptedPairs.

New finding

• [low] taste: the Selected badge squeezes the selected round's title onto two lines (frontend/src/styles.css:102). Once the row hits its 196px minimum, the selected item reads "Round / 3" and "HTML / page": its title is 48px tall where every other title is 24px. Fix: give .round-labels flex-basis:100% at all widths, as the rule for screens up to 1000px already does, so the badge drops under the title.

Requirements

• Design tokens, type and spacing: met. The palette, 36/28 title, 44px targets, radii and 1200px max width are in the tokens and styles.css. taste@1 found no contrast, motion or chrome failures.
• Fonts: met through the disclosed local fallback. The probe measured Arial in use.
• Sequence-first layout: met. Goal → sequence → large selected output → "What comes next?"; evidence sits in disclosures.
• Browse without a wallet: met. "Connect wallet" is in the header, and a hint explains browsing without one.
• Continue opens preparation only: met. Proposal, approval and payment are separate steps in ProposeForm and FundingSheet.
• Historical Fork beside the selected output: met (ArtifactViewer children).
• Latest vs Selected: met. Older rounds show a read-only note with "Select latest".
• All eligible proposals reachable: met. The Contributions list shows every proposal, each current one with "Review funding".
• Mixed-read recovery with read-only Retry: met (code plus the re