composer-chat-18: response fork verification
Base: 2b5a87909883b1abd4732b68630eb6dcef789344
Server layer: 549047dfb54a905b6d85940507cd9469735cc5b5
Live production capture head: d8dbb784d0fc1d44038e058608bd4cc7abd54c72
Final integrated head: 3b4b62cba56d8315081566b00a8f51e686be7895

Before: 318 existing tests passed in five task suites. Minimal new behavior cases produced 319 passed / 13 expected failures (332 total). After the initial implementation: 332 passed.
Independent round-1 reviews found two distinct ancestry bugs. Fresh regression stage: 30 baseline passed -> 29 passed / 2 expected failures (31 total). After fixes: 31 passed.
Independent round-2 Opus review found that rolling back the source after a fork strands its first continuation. Fresh regression: 12 baseline pass -> 12 pass / 1 expected failure (13 total), then 13 pass. The fix reads the target fork inherited projection, retaining ancestor and cut history. Both independent round-3 full-diff reviews found no blocking code findings; Opus independently reran 334 task tests. Source rollback is verified with persisted execution tests, not a live filesystem rollback.
Final focused verification: 454 passed in ten suites, comprising 334 task tests in five files and 120 existing supporting tests in five files. All five scoped package compilers passed across the implementation stages; server and shared runtime rerun after the ancestry fixes. Server compiler reran after the round-3 portable-context fix. Targeted lint passed with no new warnings, formatting and diff checks passed. No repo-wide checks.

Live Linux web / actual Codex provider, integrated head d8dbb784:
- Original source has BEFORE_BOUNDARY_18, then a real accepted same-run steer with AFTER_BOUNDARY_c18_a9f3, then its later reply. Actual base 2b5a879 has no fork action on the earlier response. Current head offers it.
- Forking that earlier response creates a visible history containing the original prompt and BEFORE_BOUNDARY_18, excluding the steer and later reply.
- Continue that fork with a new local run: LOCAL_REPLY_BOUNDARY_18, watched-file barrier, real same-run steer LOCAL_AFTER_c18_b72f, later reply. Cut at LOCAL_REPLY_BOUNDARY_18 into a grandchild and continue it without supplying marker values in the verification question.
- Current-head fresh grandchild c693b10b-a4a6-4e3f-8940-fd185182d93d continued the existing real-steer source. Actual Codex answer: BEFORE_BOUNDARY_18 LOCAL_REPLY_BOUNDARY_18 AFTER_MARKER_ABSENT.
- Persisted portable handoff delivery.status = injected. It includes both original and local pre-cut markers; neither private post-cut marker appears. This verifies provider-process delivery, not only prepared payload.
- Archive the middle ancestor, fully reload its active child, and fork the inherited root response again. Prior head c33a41 widens it to a whole run and restores AFTER_BOUNDARY_c18_a9f3. Current head d8dbb784 stores throughTurnItemId and excludes that marker.
- A real Codex stream was stopped after 164 of 2500 requested numbers. The run is interrupted; current head forks its partial response. The Codex adapter records that assistant item as completed on Stop, so this is not live proof of assistant-item status interrupted. Deterministic execution tests cover an interrupted assistant item.

Limits: Linux web and Codex only. No native Electron shell, macOS, Windows, iOS or Android runtime check. T3 device access is disabled. Other providers and remote/relay/tunnel connections were not exercised live. Base/head browser captures use the same sandbox FontFace provision because Chromium initially rendered text without glyphs; no product font change. Old clients cannot express an item boundary and can still explicitly request whole-run forks; upgrade clients to retain cuts. New clients guard old destination servers with the advertised capability.

Artifacts contain synthetic test prompts and marker values only. No production database, auth tokens, pairing URLs, or customer history is included.

Current-head recaptured root fork, fully reloaded archived-ancestor fork, Stop fork and actual continuation. Current lineage JSON and injected five-item handoff retained locally. Portable whole-run handoffs also use the target inherited timeline: ancestor history and later-rolled-back source runs remain included, superseded interrupts remain hidden; native whole-run forks are unchanged. Response item cuts carry text context without native tool/reasoning state.

CI follow-up: the existing completed-response browser fixture routed thread-N while its local item explicitly belonged to thread-1. Matching that case to its owning thread preserved the whole-run assertion. Configured Chromium file: 64 pass / 1 expected failure before, 65 pass after. Web compiler, targeted lint, formatting and diff checks passed. Final upper commit 3b4b62c changes only that fixture line; production diff SHA-256 is identical to d8dbb78: e121ec0c5826eebf7b990685e694b7ff7e4f36056777ade8fb8de3758a5007c0. Live captures remain valid and retain their actual capture-head labels. Four independent parallel full-diff review rounds completed; both round-4 reviewers found no blocking code findings and adjudicated the optimistic-steer bot comment as a false positive. Final scoped totals: 454 focused tests across ten files plus 65 configured Chromium tests in one additional file.