Complete the .18 pin: re-extract the app-bundle overlay + re-record the service-contract fixtures (unblocks #622 and main)

[aarontrowbridge]

Important

Problem — main is red on two gates, both from the same root: the .18 pin moved but its dependent artifact families never followed. #634 pinned the vendored binary to v1.18.10-amicode.18 in opencode.lock.json only — the app-bundle overlay/manifest is still pinned at .16 (with #620's hand-edited titlebar files, unreachable-green by the drift gate's design), and the service-contract golden fixtures still record .17 (the failing assertion's own message names the remedy: "re-record on pin bumps").
Approach — Complete the pin sweep mechanically: re-run the overlay extractor at .18 (overlay + manifest cohere with the pin; this carries the fork's own merged work between .16→.18, which is what re-pinning means), and re-record the contract fixtures against the .18 fork binary. This supersedes #623: the pin decision #623 held for an owner is now MADE (#634), so this is mechanical completion of an expressed decision, not a new choice.
Scope — in: extractor run at .18, fixture re-record, whatever the regenerated artifacts require · out: any code changes, any new pin decision, the TUI or deployment surfaces.
Assumptions — The .18 fork tree contains the titlebar change #620 hand-carried (verified: the .18 tag's titlebar-channel.test.ts carries the dev→DEV fix), so the extractor reproduces the hand-edits byte-identically and the drift resolves rather than reverting them.

Acceptance Criteria

• The overlay + manifest are regenerated at v1.18.10-amicode.18 by the extractor (no hand-edits anywhere in the diff), and drift_gate.mjs exits 0 with a fork clone present (full re-derivation, not just the committed-state check).
• The service-contract golden fixtures are re-recorded from the .18 fork binary and the self-tracking pin assertion passes (fixtures were recorded from the CURRENT fork pin).
• The full extension vitest suite passes; main green on both app-bundle-gate and fast.
• The parked ADR PR docs(adr): Harness Contract v1 — the fixture-pinned opencode API subset (A2, #621) #622 is then mergeable (rebased if needed) — this repair is its unblock.

Testing Decisions

The drift gate script itself (with --fork pointing at the local clone) is the overlay's test; the extension vitest contract suite is the fixtures' test; both must run green locally before the PR. No new test surface — this slice's output is regenerated artifacts.

Key Decisions

1. Re-pin to .18 — not a choice, the completion of chore: pin opencode v1.18.10-amicode.18 #634's choice. The .16→.18 fork delta (the fork's own merged PRs) rides in the overlay; that is the semantics of a pin bump.
2. Fixtures re-record from the vendored .18 binary per record_amicode_fixtures.mjs's existing flow — no recorder changes.
3. No code changes in this slice. If the extractor or recorder surfaces a real incompatibility, that's a follow-up issue, not a drive-by fix.

Constraints & Invariants

• The overlay is extractor-written only — no hand-edits in the diff.
• opencode.lock.json stays at chore: pin opencode v1.18.10-amicode.18 #634's .18 (already correct).
• The kerr_*.md untracked scratch files in the shared checkout are untouched.

Prior Art

• packages/app-bundle/scripts/{extract_overlay,drift_gate}.mjs — the sync mechanism and its gate.
• packages/extension/scripts/record_amicode_fixtures.mjs + test/amicode_service_contract.test.ts — the fixture record/replay pair with the self-tracking pin assertion.
• amicode app-bundle drift on main: #620's hand-cherry-picked titlebar overlay is unreachable-green — pick the pin and re-extract #623 — the diagnosis this slice executes (its "option A" branch, now decided).

Source

• Found while unblocking the campaign: ADR PR docs(adr): Harness Contract v1 — the fixture-pinned opencode API subset (A2, #621) #622 and the planned liveness slice both park on this red main.
• Campaign ledger sessions/session-20260827-harness-agnostic.md (personal vault), loop 5 prelude.

Notes

• The extension suite is otherwise healthy (1704 passed, 1 failed — the pin assertion), so the blast radius is the two artifact families, not the code.

[aarontrowbridge]

Scope amendment (honest record): the sweep surfaced a necessary follow-through the issue scoped out — six extension tests pinned the .16 overlay's API, which the fork's .18 deliberately reworked:

• vaultNodeLocked (context-tree-data) → became the injected opts.vaultLocked(mount) callback, covered by the fork's own context-tree-data.test.ts; effectiveMount (vault-browser-model) → inlined into vault-browser.tsx's mount memo. Four tests retired with a pointer; the two living-API tests (vaultRefFromPath + the end-to-end render) stay.
• amicode#323 (dblclick-to-close) is superseded by the fork's .18 tab UX: dblclick now opens rename, middle-click (onAuxClick + MIDDLE_MOUSE_BUTTON → closeTab) closes, and the × button stays. The regression test is rewritten to pin the current contract.

No production code changes — the sweep's diff is regenerated artifacts plus these six test adaptations. Local evidence: the three target files green (79 tests); terminal.test.ts ×2 and pasqal_python ×1 fail identically on pristine main on this machine (environment: this box's python lacks venv/ensurepip; the terminal test is HOME-dependent) and pass on CI — they are not part of this diff.