Stronghold split gesture reviews: offline repair, replay stopped at early review fetch
The two personal review waits formerly inside route-gesture are now separate dependent commands. Offline controls pass, including freshness at the eventual physical press. The sole original replay completed and inspected its separate ChooseType review in time, then stopped during hover review because the operator fetched the request before the new command published it. The capture-bound guard correctly rejected the previous command's request. No Custom Game press, later-layout qualification, preset, native Start, world or gameplay followed. This run does not establish live success for the complete repaired gesture.
The predecessor is bounded route continuation, published 03131dd67; the coordinator verified its349 files /17167938 bytes before assignment. Run-owned helpers, tests and evidence are in scratch/runs/20261104-stronghold-split-gesture-reviews/. Only this note is committed. Production code and WAT are unchanged; no local heavy work or full build was performed.
route-review-choose captures original ChooseType, requires its fresh separate personal VIEW and exact review, and checks freshness again before returning. route-review-hover checks that same presentation identity before moving, drains the owned move, captures the recognized Custom Game hover, requires its own separate personal VIEW/review, then checks freshness before returning. route-gesture accepts only that completed phase, captures again and compares against the reviewed hover identity before the balanced physical down/up. It contains no review wait. These commands retain one PhysicalRoute instance, one immutable evidence queue, and the same aggregate counters.
Source inspection establishes that freshness is transitive from the personally reviewed capture to the post-review capture and then to the capture immediately before input. assertFresh hashes every RGBA pixel in the measured output rectangle plus frozen tick, HWNDs, DOM object/session identity and geometry, checks the captures' stored presentation hashes, and requires equality. Before the move and press, exact scene recognition and drained input telemetry remain required. A failed phase terminally rejects further route commands. Later preset routing retains its original physical driver and review guards.
The unchanged limits are3072 cumulative arrival steps,768 stability steps,4096 total route steps,64 per ACK,16384 navigation steps and360000ms browser duration. Each command retains its original absolute40-second deadline, including actual reply inspection; no late success or review clears the terminal latch. Launch additionally requires a full six-minute browser window before the assigned hard close, ten minutes of remaining runtime TTL and the existing five-minute expiry reserve. No deadline extension or counter reset was introduced.
New offline checks passed six split cases: early press, early hover, foreign review hash, stale ChooseType before move, stale hover before press, and successful separate reviews with no further review wait in the gesture. The successful fixture also rejects replayed phases. Retained checks passed nine bounded-route cases,18 operator controls, six actual-wrapper personal-review controls, four actual transport cap/deadline/latch controls, five launch-order checks, and three positive/48 negative layout cases. A real wall-clock test waited40.000244 seconds and rejected an already successful reply inspected after the deadline, allowing quit NEXT only. A separate process verified the actual nonblocking exclusive operator lock. These are offline scheduling/guard fixtures, not permission for live input.
Before Chrome, actual wrapper top-level setup, CJS/ESM/Puppeteer imports, complete transitive helper SHA closure, HTTP isolation and98 page/12 Worker cache entries passed. All2415 retained runtime pins passed before launch and again after closure. Original source remains 1188e9bf57ef35bd3363fb88b7ca0da442ce696d; canonical module remains cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee, companion 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3. No latest-main module or environment rebuild was used. Preflight executed no synthetic guest callback fixture.
The sole Chrome21497 recorded identity at11:51:59.190Z and absolute browser deadline11:57:58.087Z, within the assignment's12:02 hard close. Five separately completed and inspected input-free commands reached10240 steps. Command5 separately reviewed the title and main menu before/after balanced Enter; command6 separately reviewed main menu and Crusader hover before the balanced physical Crusader gesture. Command7 freshly reviewed original ChooseType at10560 steps and installed the route driver.
Command8 route-review-choose requested a new exact personal review at11:54:10.871Z. The operator separately viewed its capture, acknowledged at11:54:19.724Z, and inspected its successful reply (written11:54:20.254Z) before original reply deadline11:54:50.367Z. No input or route steps occurred in that command.
Command9 route-review-hover began at11:54:24.900Z. The operator requested its review artifact too early: review-required.json still described command8, so review_bound rejected the command-index mismatch. No stale acknowledgement was written and no permission was granted. The live command performed only the already reviewed move, reaching10624 steps, and published its fresh hover request at11:54:26.208Z. Under the first-failure rule the operator did not fetch or review it again; instead it persisted the terminal latch. The wrapper observed that latch and rejected command9 at11:54:36.429Z, closing Chrome normally at11:54:36.710Z, exit0, elapsed157520ms from identity.
The first local attempt to send quit encountered the still-held exclusive command lock and submitted nothing. After command9's rejected reply was inspected, quit10 was the next mailbox command submitted. The wrapper had already closed, so quit10 was not processed or acknowledged; its own missing-reply inspection later timed out. No further browser or input command was attempted. There was no retry, reload, second Chrome, forced callback, register/argument write, repaint or fault clear.
actual-result.json checks166 complete64-step ACKs /10624 steps,1032 served-pin matches, six paused-before-start Worker preloads, six separate personal reviews, successful timely inspection of commands0–8, and fully drained final input. All229 immutable closed artifacts /14564074 bytes were recovered and fully SHA verified. Unique driver log, executed wrapper/route helper, release, terminal latch and command9/quit10 requests were recovered separately. Executed wrapper and route bytes exactly match the tested local files. At11:55:19Z there was no Chrome or driver and HTTP46855 was closed.
Exact Git SHA, relevant binary diff and small run-owned helper patches preserve provenance without source archives, module copies or repeated game media. Custody returned on the messageboard: bx_dkj38gmx remains RUNNING (API idle, health ok, hydrated), expiry 2026-11-04T12:13:15.406Z. The coordinator owns shutdown at12:07; this worker took no lifecycle action. The concrete remaining replay blocker is operator scheduling of the review-request fetch relative to publication of the current command's request. A separately authorized replay must wait for that current request without granting permission from the retained previous one. This single assignment ends after its note commit and one strict pull/push attempt.