Local-only follow-ups after #2253 / #2266 / #2268 — browser gates, orphaned commits, dead reconcile branches, Dev Drive

[BigSimmo]

Task for a local Claude Code session on the Windows workstation. Everything here is blocked by the cloud container specifically, not by approval or cost. Written 2026-08-22 from a web session; main at dea32b8d. Every claim below was verified in that session unless it says otherwise.

Work A1 → A4 in order. Part B is out of scope for this issue.

---

Context — what just landed, so it isn't redone

• feat(dictionary): merge Search and Browse into one catalogue and rebuild the phone header #2253 — Dictionary Search and Browse merged into one catalogue at /dictionary/search; /dictionary/browse redirects; phone control row rebuilt; plus the fix for the mobile-/ layout shift (0cf0493).
• docs(issues): record the mobile-/ CLS root cause and file the hydration-gap guard #2266 — root-cause record for that layout shift, written against #TYZK23, #50QRCF, #KFRC3H, plus new #2TAQDC.
• docs(issues): unstick the ledger write path — reconcile 21 queued requests and rescue two orphaned review records #2268 — the outstanding-issues write path unstuck: 21 queued requests reconciled, two orphaned branch-review records rescued.

Ledger is current as of dea32b8d. Pending inbox is zero.

---

A1 — Recover or discard the two commits that exist only on the workstation

• Decided, and said which — cherry-picked as PR design(dictionary): measure the control row's intrinsic width in the fit strip, not its laid-out width #2281

Why local-only: these objects are not on origin and are unreachable from any container. Verified: git cat-file -t 970d28d25 → Not a valid object name; git ls-remote --heads origin claude/phone-dictionary-header-9180ee returns nothing.

Per the original Dictionary task brief: 970d28d25 fixes the mockup's fit-measurement code to probe intrinsic width; e8e7371b5 merges origin/main. Branch claude/phone-dictionary-header-9180ee.

git log --oneline -5 claude/phone-dictionary-header-9180ee
git show --stat 970d28d25

Then either cherry-pick onto a fresh branch off origin/main and open a PR — mockups are exempt from the wiring and reachability gates but are still typechecked and still weighed by check:bundle-budget against the mockups baseline, so run npm run verify:pr-local — or state explicitly that it no longer applies and delete the branch. Do not silently drop it; this is the last unresolved item from that task brief.

Related and still open: the Codex review comment on #2241 (comment_id=3831169288) that 970d28d25 was written to address.

A2 — Run the browser gates no cloud session has been able to run

• verify:phone-chrome green — 133 contract tests passed, focused browser 12 passed
• verify:ui green — 473 passed, 2 skipped, 0 failed (19.7m)

Why local-only: the container has Chromium 1194 while Playwright 1.62.1 pins 1234, and the network policy blocks cdn.playwright.dev, so npx playwright install chromium returns 403. Verified this session:

Playwright browser revision check FAILED (not-installed): Playwright expects chromium
revision 1234, but no matching install was found at /opt/pw-browsers
(found instead: chromium-1194)

#2253 therefore shipped without the full Chromium gate. Two suites ran through the config's documented PLAYWRIGHT_CHROMIUM_EXECUTABLE_PATH escape hatch against the mismatched revision (ui-dictionary 7 passed, ui-mode-nav-density 48 passed) and CI Production UI passed — but verify:ui in full and verify:phone-chrome have never run against the merged Dictionary work.

npm ci --include=dev
npm run check:installed-lock-parity
npm run check:playwright-browser-revision # must report OK before the rest means anything
npm run ensure # use the URL it prints; never assume localhost:3000
npm run verify:phone-chrome
npm run verify:ui

Report the decisive "N passed" line, not the exit code. verify:ui exits 75 with a DATABASE_HEAVY_RUN_ADMISSION_BUSY marker under lock contention — that is "blocked, retry", not "red" — and exit 0 alone is not proof.

Measurement trap #XPY409: .phone-sticky-header-stack is position: fixed and mounts collapsed. Wait ~1200–1400 ms for settled height before measuring or screenshotting phone chrome.

A failure here is a real regression in merged work. Fix it with the smallest change and open a PR; do not weaken the test. In particular do not drop production tap targets from min-h-12 to min-h-11 — that reintroduces a known ui-smoke flake and is forbidden in AGENTS.md.

A3 — Delete the four dead reconcile branches

• All four deleted, and issues:reconcile --dry-run no longer refuses

Why it matters: they blocked ledger reconciliation for five days and will block the next one identically. assertSafeRemoteReconciliation refuses to start while a branch whose name indicates an in-flight reconcile exists on origin (#EH9VA6).

Already established in #2268 — no need to re-derive. For each of the four, every docs/outstanding-issues-inbox request UUID on the branch is already on main, in either the pending or applied/ directory, so nothing queued is lost. Their only content absent from main was two branch-review records, rescued in #2268. The remaining branch-only files are five sources main deleted deliberately in #2211 (use-upload-desktop-layout.ts, service-group-nav.tsx, therapy-compass/icons.tsx, two upload tests). None had an open PR as of 2026-08-22.

git fetch --prune origin
for b in claude/issues-reconcile-2026-08-19 \
 claude/ledger-reconcile-2026-08-21 \
 claude/ledger-reconcile-issues-c1trbj \
 claude/rag-r0-reconcile-inbox; do
 gh pr list --head "$b" --state open # must be empty for each
done

for b in claude/issues-reconcile-2026-08-19 \
 claude/ledger-reconcile-2026-08-21 \
 claude/ledger-reconcile-issues-c1trbj \
 claude/rag-r0-reconcile-inbox; do
 git push origin --delete "$b"
done

npm run issues:reconcile -- --dry-run # must NOT print the unmerged-branch refusal

Deletion is irreversible. If any of the four has grown an open PR since 2026-08-22, stop and report instead of deleting.

A4 — Reclaim the Dev Drive (#6GW95D, P3)

• Inventory taken (208 checkouts across 8 roots) — removal pass halted and fully reverted by owner decision; cleanup deferred

The ledger's own update is the point: the scope is far wider than .claude/worktrees. An accidental npx vitest run from C:/Users/joshs enumerated 253 copies of just two test files across the home directory — 253 separate checkouts of this repository across at least six roots (.codex/worktrees, .gemini/antigravity/worktrees, .copilot/repos/copilot-worktrees, .local/share/opencode/worktree, Documents/Codex/..., and more). The capacity emergency is resolved; the inventory and safe orphan cleanup are not.

Inventory first, delete second. Only remove a checkout with no unpushed commits and no uncommitted changes. Read docs/branch-cleanup-guide.md first. Treat any checkout you cannot prove is clean as one to keep.

---

Part B — explicitly NOT this session

Each is possible on the workstation but is its own piece of work with its own approval. Tracked already in docs/outstanding-issues.md; listed here only so the boundary is unambiguous.

Item	Why it is separate
#1YPV51 (P1)	The medication interaction lexicon is clinically unreviewed despite #318 recording otherwise. Owner's clinical judgement — no agent does this.
#1K6T35 (P2)	Point-in-time recovery is OFF on the live project (pitr_enabled false, measured 2026-08-19). RPO up to ~24 h on 2,851 documents / 70,120 chunks, and it blocks the "restore point before any mutating phase" rule that #056/#057 depend on. Dashboard/billing decision.
#CCZ4HB (P1)	Review-bot spending cap reached; three PRs merged 2026-08-22 with no automated review. Billing decision.
#056, #057, #102	Supabase windows. check:drift is red with 19 findings, re-measured unchanged 2026-08-18; schema.sql is the stale side for at least seven. Merging under supabase/migrations/** deploys to the live clinical database within seconds.
#001, #231, #018	Provider-backed live evaluation; costs money and needs explicit approval.
#S4K1GA, #71NT23, #75JA0P	Physical iPhone Safari and installed-PWA acceptance. Chromium and Playwright WebKit are not the iOS engine.
#QSHHGK (P2)	Bundle-budget baseline has no owner or trigger. Note: npm run build reuses a cached .next, so rm -rf .next before measuring or the number is stale.

---

Standing rules

• Never report a gate as passing that was not run. Paste the decisive line; exit 0 alone is not proof, and a reused gate receipt is reported as "reused receipt", never as a fresh run.
• Providers need explicit confirmation — OpenAI, Supabase, GitHub, hosted CI — including indirectly through scripts and release gates.
• Flag the task before editing anything under src/lib/rag/**, clinical-search, retrieval-selection, ranking-config, answer-ranking, the eval harness or the golden fixture, even for a rename.
• Run npm run format and commit the result before any normal push. It is in neither test, typecheck nor lint.
• Never assume localhost:3000. Use npm run ensure.
• The ledger is append-only from feature branches: npm run issues:add|update|done; only a dedicated fresh-base branch runs npm run issues:reconcile.

[BigSimmo]

A3 update — attempted from the cloud container on 2026-08-22 and refused.

Branch deletion is not possible from a Claude Code web container. The git proxy returns 403 on a delete refspec:

$ git push origin ":refs/heads/claude/issues-reconcile-2026-08-19"
error: RPC failed; HTTP 403 curl 22 The requested URL returned error: 403
send-pack: unexpected disconnect while reading sideband packet
fatal: the remote end hung up unexpectedly

Note the trap: the plain git push origin --delete <branch> form reports "Everything up-to-date" and exits without an obvious error, so it looks like it worked. It did not. Always verify with git ls-remote --heads origin afterwards rather than trusting the push output.

The GitHub MCP server exposes create_branch but no delete-branch tool, so there is no API route from a container either.

Pre-checks that were completed and still hold (re-confirmed 2026-08-22 against the full open-PR list): none of the four branches has an open pull request. All four are still present on origin at d2120b2, 110db29, d439b65 and 6f5503f respectively.

Two ways to finish A3, both outside a cloud session:

1. GitHub web UI — repository → Branches → delete each of the four. Roughly thirty seconds, no local checkout needed.
2. Local session — the commands already in section A3 above work fine from an authenticated workstation.

Either way, finish with the evidence command:

git fetch --prune origin
npm run issues:reconcile -- --dry-run # must NOT print the unmerged-branch refusal

---

Generated by Claude Code

[BigSimmo]

Part A complete — all four items, with evidence

Run on the Windows workstation, 2026-08-22. Part B untouched.

A1 — the two workstation-only commits: cherry-picked, merged

Both objects were present locally on claude/phone-dictionary-header-9180ee. The task brief's framing needed correcting, though: the Codex finding on #2241 (3831169288) had already been fixed twice, independently. A fix was applied on the PR itself and squashed into main as 6b4ea9bfd (a data-fit-text marker plus adding the clipped scrollWidth - clientWidth back into the summed width). The unpushed 970d28d25 solves the same finding differently — it probes the whole track under max-content for one synchronous read, so every child sizes to its own content regardless of flex-1 or ml-auto, and reports truncation as its own figure instead of folding it into the slack number.

So this was a choice between two live fixes, not a recovery. Decision: take 970d28d25, because it measures the row's intrinsic width rather than patching the arithmetic, and because the separate truncation figure is the measurement the design conclusion rested on. The cherry-pick conflicted, as main's version of that function is no longer what the commit was written against; resolved by taking the max-content probe wholesale (which makes the textOverflow add-back dead code) while keeping main's data-fit-text marker name, so the diff is the measurement change and nothing else.

Landed as PR #2281 and verified on main by content, not by PR state — origin/main now carries the max-content probe and queryShortfall, and no longer carries textOverflow.

No automated gate can observe a measurement change (test:e2e:pr runs --grep-invert "@mockup"), so it was read live in Chromium at the URL npm run ensure printed:

Measuring idle 320 wraps 15px short | 390 one row 55px spare | 430 one row 95px spare
Measuring with "tardive" 320 wraps 31px short, query cut by 30px | 360 one row 9px spare
Measuring with "tardive dyskinesia" 390 wraps 28px short, query cut by 28px | 430 one row 12px spare

That reproduces the original commit message's figures exactly, and it is the evidence behind giving the search term its own line: "tardive dyskinesia" does not seat on one row until 430 px, which no phone in the width table provides.

The review thread 3831169288 was found already resolved, closed by the owner when #2241's own fix landed — so it needed no reply, and per AGENTS.md a resolved thread stays resolved. The issue's "still open" note was inaccurate.

A2 — both browser gates green

check:playwright-browser-revision reported OK (installed): Chromium revision 1234 is installed and launchable, so the container's blocker does not exist here. check:installed-lock-parity passed, so no npm ci was needed.

A trap worth recording, because it nearly produced a false green. Invoked bare, verify:phone-chrome exited 0 having run no browser tests at all — its changed-file selector found only an untracked inbox JSON, printed No phone-chrome-affecting file was detected; only prerequisite checks were selected, ran two docs checks and exited clean. Reporting that as "green" would have been exactly the failure this issue's standing rules warn about. Re-run against the merged Dictionary scope (git diff --name-only 66594ddc1^1..main, filtered to src/ and tests/ui-), the selector produced the real plan and the real results:

Stage	Result
lock-parity, runtime	passed
contracts	Tests 133 passed (133)
focused-browser	12 passed (49.0s)
full-ui → verify:ui	473 passed, 2 skipped, 0 failed, (19.7m)

Exit 0, first attempt, no admission-busy retry. No regression in the merged Dictionary work. Filed separately: the selector's inability to distinguish "nothing selected" from "everything passed".

A3 — four branches deleted

All four confirmed still on origin at d2120b2b, 110db29c, d439b650, 6f5503f9 with zero open PRs, then deleted. Verified with git ls-remote --heads origin, never the push output — the comment's warning about a silent "Everything up-to-date" was well founded. No reconcile branch remains.

The refusal is gone: npm run issues:reconcile -- --dry-run from a clean checkout now prints Would reconcile 5 request(s) into docs/outstanding-issues.md. and exits 0. (Run from the main checkout it reports a different refusal, about an untracked pending request there — unrelated to this item.)

A4 — inventory complete; cleanup halted and fully reverted

The inventory supersedes the 253 estimate with a measured figure. Eight roots, 208 checkouts of this repository: 92 registered worktrees of D:/Repos/Database and 116 separate full clones with their own object databases (76 under .codex, 26 under .copilot, 13 under Documents/Codex, 1 under Temp). Zero unregistered worktrees, zero stale gitdir pointers. clean:worktree --merged --squashed --dry-run found 23 landed candidates totalling 9.30 GB. Capacity is no longer the driver — D: is 80 GB with 53 GB free.

The single most useful finding: scripts/clean-worktree.mjs cannot see the 116 clones at all. They are the real disk mass, and no worktree-scoped sweep will ever touch them.

12 contentless leftover directories were removed. The removal pass over landed worktrees was then halted by the owner and fully reverted: all 13 removed worktrees restored, each verified individually for correct branch, correct head and clean tree, with dependencies restored by byte-identical copy rather than reinstall. Five live chats had had their working directory deleted underneath them; all five work again. No commit, branch or uncommitted change was lost at any point — git worktree remove never ran with --force and refuses a dirty tree.

Two process lessons are recorded on #6GW95D, both learned the hard way here: a git refused / EPERM removal means a live process holds that directory and is a stop, not something to retry on a later pass; and the candidate list must be shown and approved before any deletion. Cleanup is deferred indefinitely.

Filed for follow-up — PR #2284

Four new findings plus the #6GW95D update, as immutable inbox requests: the verify:phone-chrome false-green above (P2); two tests/gate-receipts.test.ts chmod assertions that cannot pass on this ReFS Dev Drive, leaving every local npm run test two-red and stopping verify:pr-local before build (P3); the Claude push-format-guard hook blocking all pushes from linked worktrees by comparing core.hooksPath against the worktree root instead of the common git dir (P3); and seven registered worktree directories that are empty while still being the cwd of seven live sessions (P3, pre-existing).

---

Generated by Claude Code