Aranna default pager: rendered menu, unresolved activation and commit refusal

Task NEW-GAME-ARANNA-DEMO-20261010; worker aranna-default-pager2220. Evidence: scratch/runs/20261105-aranna-default-pager-gameplay/ (ignored).

The sole repaired original passed the normal EULA, rendered an opening cinematic, and reached the Main Menu after ordinary Escape. Single Player clicks, a held click, Enter, and held Enter did not reach character setup or ordinary gameplay. The menu continued rendering. Later console output records an explicit heap boundary: a 1 MiB sparse reservation could not be committed, followed by smaller heap requests returning failure. No additional original was launched. FPS and audio are unknown; two gameplay responses were not reached.

Generated real PE controls prove the published 2915073f6 default MEM_RESERVE repair still enters guest SEH, commits the reserved page, performs a real nested ReadFile, restores the interrupted CPU, and resumes reading the filled value in both production owning-Worker module variants. The original's passive trace exhausted its 4,096-record cap during earlier unrelated checksum calls. Consequently this run does not directly prove the original's pager entry, commit/read/resume, or resumed Logic.dsres validator. Menu progress is a separate observed outcome, not a substitute for those missing observations. There is no demonstrated new emulator defect and no speculative source repair.

Gate and source identity

SOURCEONLY ended after the exact FINAL releases, in order: Conan world writer at 22:51:21.579564Z, D3D worker reset at 22:59:46.472632Z, and AoW2 black ground at 23:13:08Z. Independent cleanup verified 996 absent PIDs, 237 refused ports, 86 absent profiles, four free locks, no heavy runtime, and sufficient disk. The immutable execution window began 2026-11-05 23:14:00.327444Z, deadline 23:59:00.327444Z. release-gate.json and runtime-bound.json retain the checks and times.

Runtime Git SHA: ed024097db9e3d51eb04dff64d32d3563b0b45dc, including ancestor 2915073f6. runtime-source.json, validation-source.json, and before-original.json identify the relevant files. Their .binarydiff files and relevant untracked-file patches preserve non-Git hunks; owned helpers are retained as new-file patches. The 319 tracked relevant source hashes stayed unchanged through controls, required full build, and original launch. No shared CPU/Win32/memory, renderer, launch mapping, or test source was edited by this worker. The only committed change is this findings note.

The original's actual browser-owned module was 2,176,244 bytes, SHA-256 681c4d52e019edce8fbcee1413e07637cd855baf8b69e168a2ff6b5c0535dc27. browser-modules.json, the uniquely named retained module, worker-start records in observations.jsonl, and served-responses.jsonl identify what actually ran. This diagnostic module differs from the canonical build only by the retained passive compiler transform; it contains no fixture export. launch.json also records a canonical build hash: that field alone is not the identity of the overridden diagnostic module used by the owning Worker.

All 28 canonical media inputs were hashed in place under scratch/game-media/aranna-demo-20261105/extracted/; see media-verification.json. The existing normal original-local app mapping was referenced in place. No download, extraction, media copy, source archive, subworker, box, public deployment, or unrelated game was used.

Validation before the original

browser-controls.json and browser-controls.log pass the generated real PE, normal DLL initialization, actual owning-Worker controls for standard and compat modules. API return values, last error, ABI/nonvolatile registers, ordinary/chained consumers, and lazy flags agree with the baseline.

pager-controls/results.json and pager-controls.log pass all 16 combinations of standard/compat, baseline/passive, already committed/reserved, and eager/ provider-backed lazy file access. The ordinary generated EXE installs FS:[0], reserves 113,820 bytes, and reads it. Its own handler commits a page and uses ReadFile to fill four bytes. Reserved cases call the handler once, return BOOL=1/count=4 from ReadFile, resume with 0x12345678, and restore all copied registers. Committed cases do not invoke the handler. Copied default mode is zero; no fault-mode setter is used. Passive fault events are [5,6,7,8] for eager data and [5,6,9,7,8] for synchronous provider fill. Both owning-Worker module identities and normal DLL DllMain EAX=1 are asserted.

flock /tmp/wine-build.lock bash tools/build.sh passed; see build.log. Canonical output sizes were 2,175,008 standard and 2,177,654 compat bytes. Integration runs were serialized; the full build completed before the sole original. An initial shell wrapper failed after six module compiles, before controls: editing its own ownership label while bash was reading it disturbed parsing. validation-wrapper-failure.json and validation.log retain this operator error. The continuation separately completed both control suites and the locked build successfully; no original was consumed by that correction.

The one original and reviewed ordinary input

The session launched at 23:17:01.672Z, using the production owning Workers, normal mss32 initialization, EULA, threads/WebGL, existing eager-file setup, and Puppeteer defaultViewport:null. startup.png was reviewed before the ordinary EULA Accept click. after-accept.png showed the ordinary loading window; loading-30s.png showed the rendered opening cinematic. Ordinary Escape produced the Main Menu in after-escape.png, with version 1.50 and Single Player, Multiplayer, Options, Continue, About and Exit controls.

Fresh captures were reviewed before further input. The Single Player click, Enter, a trusted mouse click held 1,000 ms through the existing Chrome, and Enter held 750 ms left the main menu displayed. Commands, completed command records, ordinary-inputs.jsonl, captures and corresponding .state.json files retain the exact sequence. No character setup, accepted Single Player transition, or player movement was observed. Logged WM_LBUTTONDOWN/UP events name the main window; keyboard messages name hwnd=0. A later copied host mouse position was centered. Those observations do not establish a causal input bug.

Read-only completed renderer snapshots at 23:22:30.071Z and 23:22:32.479Z show root owning endpoint 1, legacy WebGL, slice sequence 26093 → 26178, completed fence 27367 → 27765, draws 127435 → 129465 and publications 1934 → 1963. Errors, fallbacks, and software completed draw counters were zero at those landmarks. renderer-progress.json retains the full snapshots. These counters prove menu rendering progress, not gameplay or measured FPS; child renderer ownership is not attested by this capture tool.

The final menu-oom-boundary.png and terminal.png were personally reviewed: main menu controls remain rendered, without the earlier Logic.dsres error17 dialog. Earlier review wording about a degrading background was too strong; reviewed-terminal-correction.json records that differing background imagery has no controlled visual oracle or proven causal link to allocation failure. The ordinary quit completed; Chrome exited 0 at 23:24:40.336Z and the probe exited 0 at 23:24:40.409Z. Owned local runtime was closed and cleaned; final-cleanup.json records independent process/port/profile/lock checks. No boxes were assigned: box ID/state/expiry are N/A.

Exact remaining boundaries and diagnostic repair

The first unresolved ordinary transition is activation of Single Player from the main menu. Later, console.log line 127 is the first explicit memory refusal: 1048576 bytes (0x100000) with reason 3, sparse reserved range could not be committed. It is followed by reason 1 for a 34,408-byte allocation and similar requests. There are 2,538 heap refusal lines by clean closure. remaining-boundary.json records the first line, sizes and limitations. Raw console lines have no timestamps, so the exact first refusal time is unknown; the host UI log does not contain these messages and cannot timestamp them.

Source inspection places reason 3 in $heap_sparse_alloc in src/10-helpers.wat: reservation succeeded, $virtual_map_commit failed, and reclaim/retry either found no releasable arena or also failed before the reservation was released and NULL returned. $virtual_map_commit_locked has multiple refusal causes, including lease exclusion, map-record capacity, backing placement, and PTE publication. No live map/backing/lease census was retained. This evidence does not distinguish exhaustion, fragmentation, metadata limits or a bookkeeping defect, and does not establish that heap failure caused the earlier activation failure. No memory option was changed.

The original retained exactly 4,096 snapshots: 3,093 event3 entries and 335/334/334 API entry/handler/return records, all root endpoint 1. No fault events 5–9 or Logic.dsres owning ReadFile return survived. The shared CRC entry 0x41b19c was observed for unrelated earlier callers, consuming the cap before the intended resource consumer. Missing events cannot be read as absence of guest paging or absence of a subsequent error.

After closing the original, run-owned narrow-passive-transform.cjs filters that shared CRC entry by the copied stack return 0x655594, retaining the existing resource-consumer range and nonfaulting scalar reader. It does not change shared source or the original module. Generated ordinary PE controls place the intended caller before 4,200 unrelated calls, then execute the same real reserved-page pager. The old observer fails the precise caller filter assertion (257 CRC snapshots versus 1 expected); the narrower observer passes standard/compat × eager/lazy, retains the correct target return, and preserves identical guest completion, handler count, fault address, filled value, ReadFile BOOL/count and provider-read results. See stress-before.log, stress-after.log, and observer-filter-comparison.json.

The calibration also establishes a coverage limit: only 256 of the 4,200 unrelated calls traverse this passive entry hook once code becomes hot. The generated test validates filtering and pager neutrality, not full-cap exhaustion or complete hot compiled-block coverage. The original's cap exhaustion is established by its own trace. This tested helper is available for a separately authorized materially changed investigation; it was not used to launch a second original.

Next work should discriminate actual sparse commit refusal causes and the ordinary menu-input transition using bounded owning-CPU evidence with separate budgets for pager, target validator and allocation-failure observations. Require a failing generated regression before any CPU/Win32/memory repair. Do not repeat this unchanged original or claim original paging/gameplay from the generated controls.