Stronghold corrected native start: integrated; menu route stopped
The corrected scenario guard from 6004e338eeecbdac153284cd363d499843bc68a4 is integrated into the immutable driver from eef09d6881329f916577eb6ac78e93e36f398d6d. The single browser attempt stopped at an earlier menu dependency: pressing Custom Scenarios left the main menu visible with input fully drained. The scenario Start gesture was never attempted. World entry, recruitment and combat were not reached.
Evidence: scratch/runs/20261104-stronghold-corrected-native-start/. The small run-only browser-driver.cjs, integration.test.cjs, navigation-plan.cjs and navigation-plan.test.cjs are committed alongside this note. Remaining helper recovery is recorded in owned-new-helpers.patch; dependencies are the unchanged durable native-start guard and immutable boundary, staged as native-start-boundary.cjs and evidence-boundary.cjs. No production file changed or production defect was established; no full build was required.
Integrated boundary and its limits
install-native installs NativeStartDriver and guardNativeCommands inside the existing EvidenceBoundary command queue. native-start returns its result in the final command receipt. Startup geometry, exact scenario/world pixel checks, guest (718,569) projection, per-guest step ACKs, one activation, input drainage and terminal failure remain unchanged. The helper is copied exactly, rather than weakening its predicates.
Personal review writes a pending request and waits for a separate operator acknowledgement containing the exact PNG SHA-256, scene, explicit personallyViewed:true and a written observation. It never echoes a request automatically. Fresh capture review precedes activation; the original guard recaptures to reject stale review. No such request was reached in this browser because the run stopped before installing the scenario guard.
After installation, individual raw key/mouse commands are rejected. A queued ordinary gesture first requires the guarded world capture, awaits separate personal review and recaptures before balanced movement/press/release. It checks complete ACKs while draining each phase under a finite execution budget. It preserves the original exact world recognizer; no later scene generalization was made. These paths have offline integration coverage, but were not exercised in this browser. Positive ordinary gameplay remains unproven.
First unexpected menu state
I personally reviewed fresh review/startup2048.png, loaded6144.png, title.png, main-menu.png and custom-menu.png before dependent navigation. Startup showed loading artwork, then the Firefly splash. Escape opened the title dialogue, and Start Game opened the original Demo V1.1 main menu.
I then followed the predecessor's misleading custom-menu name and pressed Custom Scenarios at browser (361,658). The deciding picture remained the main menu. At step 7936, first-unexpected-state.json proves inputQueue:[], activeInput:null, lastInput:null, mouse mask zero, and the visible original 800×600 window. This was not input still awaiting delivery. No more stepping or game input followed; only state, snapshots and quit.
After closure I personally reviewed the predecessor's custom-menu.png and choose-type.png. Its purported custom menu is also the main menu. Only its following Crusader press reaches “Choose type of Crusader game.” The original route independently names Crusader at guest (304,144). A disabled or unavailable Custom Scenarios control is consistent with the pictures; its internal guest handler was not established. I should have inspected that intermediate predecessor picture before treating its filename as a route instruction.
The offline navigation-plan.cjs now identifies the main menu and Crusader-type dialogue using exact reviewed RGB label patches. It proposes Crusader, native (304,144), projected to browser (337.3375,360.4433333333333) under the supported geometry. It rejects both historical and current unchanged “custom-menu” pictures as a successful transition. This correction was tested offline after the browser closed and is not a newly demonstrated native transition. No second browser, reload or substitute gesture was used.
Validation and evidence custody
Before browser launch, all 2415 local and remote runtime pins, locked Puppeteer 25.7.0 actual CommonJS/dynamic imports, five real loopback HTTP responses with COOP/COEP and exact hashes, the original native-start model suite and immutable snapshot suite passed. Modified-driver fixtures passed locally and remotely: actual queue serialization, snapshot/close ordering, balanced controls, capture publication, incomplete ACK rejection, locked world controls, terminal failure and a review that remains pending until explicit fixture acknowledgement.
The later navigation regression accepts the reviewed Crusader-type picture, rejects unchanged menus and a corrupt pixel, and rejects unsupported geometry. Its initial candidate native (302,144) failed the existing inverse's round-trip check; the final plan uses the original documented (304,144), which passes. Startup (718,569) and its guard were unchanged. An earlier integration assertion expected the old “locked” error after adding a stricter balanced-gesture prohibition; the updated fixture separately verifies both prohibitions. Initial failure logs are preserved. Neither was an additional runtime attempt.
Useful checks, from the repository root:
node docs/re-notes/stronghold-native-start-boundary-20261104.test.cjs
node docs/re-notes/stronghold-stable-recovery-combat-20261104.test.cjs
node scratch/runs/20261104-stronghold-corrected-native-start/integration.test.cjs
node scratch/runs/20261104-stronghold-corrected-native-start/navigation-plan.test.cjs
python3 scratch/runs/20261104-stronghold-corrected-native-start/verify.py
Atomic immutable snapshots round-tripped by full SHA-256 and length: live-a 79, final-live 296, closed 301 logical files, reusing matching retained bytes. Live downloads used published immutable manifests, never active log manifests. verify.py checks all recovered hashes, 141 sequential successful command replies, 124 step requests, maximum 64, total 7936 steps, all 1033 served-file pins, capture hashes, final input drainage and absence of gameplay after the deciding capture. Chrome 8597 ran 05:40:32.238–05:42:41.152Z, 128.914 seconds, exit zero, within the single twelve-minute bound.
Fresh minimal runtime staging used the predecessor's exact Git/blob reconstruction recipe, original installed-media path/hash references and retained canonical modules: cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee and 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3. Runtime source remains 1188e9bf57ef35bd3363fb88b7ca0da442ce696d plus the recorded nonexecuted census delta; original EXE remains 9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f. Original media was staged once onto the fresh assigned box, with no media or module copies in this run evidence. Source custody uses Git SHA, relevant binary diffs and owned new-file patches, without source archives.
Actual exec 05:43:39.640731Z proves Chrome/driver absent, private HTTP port 34289 closed and all 2415 pins unchanged. Independent API inspection reports bx_c5e39n22 RUNNING/API-idle, healthy and hydrated, expiry 2026-11-04T06:21:54.275Z. Recovery finished before 06:11Z. The box is left running; shutdown ownership returns to the coordinator. No allocation/resume/stop/delete, local heavy runtime, subworker, successor, public deployment, guest write, cheat, SaveLoad or game-check change occurred.
Next authorized native work should take the directly observed Crusader route, verify each resulting menu, then exercise the integrated single scenario Start. This task stops at the diagnosed menu-route dependency; it supplies no new world or combat claim.