Aranna: measured menu heap refusal boundary

One original run reached the EULA, rendered cinematic, and Main Menu, but did not reach Single Player setup or gameplay. The first returned heap allocation failure was an emulator temporary vertex-packing allocation, not an identified guest malloc. No production repair is justified by this evidence alone.

Evidence is retained in scratch/runs/20261105-aranna-menu-heap/ (ignored). The tested source was main 11a9dc75cf3701d66e001b1374533a81a97f7099. validation-source, controls-source, before-original, and final-source records contain exact source hashes, relevant git diff --binary, and patches for new run-owned helper files. Git is the source archive. The 28 canonical media files were referenced and rehashed in place at scratch/game-media/aranna-demo-20261105/extracted/; none was copied or downloaded.

Boundary actually observed

observations.jsonl, heap-summary.json, diagnostic module callsite tables, and console.log distinguish three events:

Four bounded censuses completed under the existing map writer lock. The first two show actual backing placement exhaustion:

Measured quantity 00:44:30.614Z 00:44:32.368Z
Sparse map records 522 524
Backing occupied bytes 301,248,512 301,350,912
Backing free bytes, total 741,376 638,976
Largest backing gap 327,680 327,680
Shared heap arenas 298 298
Arena reserved bytes 315,269,120 315,371,520
Arena live charged bytes 9,183,360 9,183,360
Arena high-water bytes 299,874,152 299,947,200
Sparse backing mapped to heap arenas 255,500,288 255,602,688
PTE/backing inconsistencies 0 0

The fixed 512 MiB module has a 301,989,888-byte sparse backing range [0x08000000,0x1a000000). Both aggregate free backing and its largest gap are smaller than a minimum 1 MiB arena. No overlapping map ranges were found. The low window also had less than a minimum arena remaining; current low and sparse bump tails were about 3 KiB. The second census follows the existing whole-free-arena reclaim/retry path; that path was not missing or bypassed. There were 982 retained decommitted pages, so these numbers describe backing placement/ownership, not simply resident committed payload.

Most arena reservation was not live charged data. This supports stranded free space/fragmentation as a substantial pressure mechanism. It does not prove a complete free-block graph, absence of a guest leak, or absence of corruption outside the checked map/PTE structures. Arena charges may race with peer frees even while map/reservation records are locked. Ordinary free lists are local to instances, while the arena table is shared. The current-instance ordinary list hit its 512-node census cap; peer lists were not walked. Both large-block allocation/free journals hit 8,192 records at 00:43:08, before the first refusal. Consequently the retained prefix cannot assign every live block at failure.

Generated comparison, without changing the original

fragment-controls.json records four ordinary owning-Worker PE cells: standard/compatibility modules, each baseline/diagnostic. All guest output words were identical. With unchanged fixed memory, each case:

  1. Allocated four 16 KiB blocks, retained a guard in the fourth, then held another 21,920 allocations until ordinary HeapAlloc returned zero.
  2. Freed the first three through HeapFree. A 45,056-byte request still failed.
  3. Called ordinary public HeapCompact, which returned 49,244 usable bytes. The same 45,056-byte request then succeeded inside the freed span, and the adjacent live guard remained 0x5a1234a5.

The initial review incorrectly assumed a 16,392-byte pointer stride. All four cells had already completed with identical results: the real stride is 16,416, because each public allocation reserves a 16-byte Toolhelp record (24 charged bytes), and HeapFree frees that record too. Source confirms this in src/09a0-handlers-base.wat:5491 and src/09a0d-toolhelp.wat:33,73. The failed assertion/log and correction are preserved; corrected assertions were run offline against retained cells, without a runtime repeat.

This demonstrates recoverable allocator fragmentation in a generated control. It does not establish that the original's complete free graph can satisfy the request after compaction, nor that implicit coalescence is a required Win32 semantic. Existing public HeapCompact already sorts/coalesces same-arena extents; ordinary allocation misses do not invoke it. No speculative global coalescing change, memory increase, guest modification, or failure suppression was made.

Input and reviewed product captures

The original used trusted ordinary mouse holds and Escape/Enter key events. Actual owning-Worker copied DispatchMessageA WndProc frames, rather than sender logs alone, establish delivery of Single Player mouse down/up and Enter down/character/up. The down coordinates were (398,152) and the up coordinates were (400,300); copied GetCursorPos outputs also changed accordingly. Cursor recentring/mapping semantics were not independently isolated, so this is not proof of correct game hit testing or an input defect. No heap/input causal claim is made from their temporal proximity.

Personally reviewed captures in the run:

There are no two gameplay response captures to review. FPS and audio are unknown.

Validation and observer limits

The locked full tools/build.sh passed, followed by focused tests: test-heap-oom-trace, test-heap-oom-stack, test-heap-arena-release, test-heap-arena-cross-instance-release, test-heap-arena-tail-reuse, test-heap-large-free-map-pressure, and test-default-reserved-worker-web. Twelve ordinary owning-Worker heap/virtual/input cells passed in both modules, baseline versus diagnostic, including exhaustion, accepted free/reuse, WndProc arguments, return/stack behavior, and nonvolatile register/branch sentinels. An initial free-journal review assumption confused an allocator-created tail with the first explicit guest free; retained cells were corrected/reviewed offline, without runtime repetition. Sixteen pager cells passed across both modules, baseline/diagnostic, committed/reserved and eager/lazy: saturated generic API observations did not consume the independent targeted validator/pager budgets, and real guest SEH committed a reserved page with nested ordinary file input.

Baseline modules were byte-identical to canonical build outputs:

Diagnostic module identities were:

The original root and seven auxiliary Worker starts all identified the standard diagnostic module above. A compiler-helper omission of the canonical region layout section was caught before execution; all four unexecuted unstamped artifacts are retained. The helper was corrected to compute the real canonical layout stamp rather than copying or bypassing one.

Observation was passive copied scalar state through the actual owning Worker; no live CPU setters, execution exports, breakpoints or guest patches were used for the original. The run transform preserved single evaluation of allocation calls and recorded exact source sites only on refusal. Census bounds, journal saturation, zero observer errors and module identities are explicit in the logs. Mirrored host-instance calls are not a complete owning-Worker journal. The targeted validator and pager budgets each reached 128 in the original; the pager prefix contains 32 complete episodes, not proof about later activity. Existing hot compiled block coverage remains limited. Observation changes timing.

The predecessor 25cd5f11331d880ea27277b8b2f0427fdf1bcee2 remains distinct: its 2,538 heap refusals, first 1 MiB/34,408-byte notifications, generic observer saturation before the target pager, and menu-only captures established neither allocator defect nor input causality. Its later filtered helper controls were not a second original. This run supplies a new measured boundary, not a retroactive interpretation of that missing evidence.

Release and next work

Source-only gating was preserved through Conan2300, legacy2305, and D3D9multiply2320 exact final releases and independently verified cleanup. The immutable execution window was 00:31:06.041430Z–01:16:06.041430Z on 2026-11-06. Exactly one materially changed original ran, 00:38:55–00:46:27, then exited normally; no unchanged original was repeated. Generated follow-up controls ended at 00:50:36. Independent cleanup at 00:52:18 verified 159 absent PIDs, 24 refused ports, 19 absent profiles, four free locks and no heavy runtime. No boxes were assigned; box ID/state/expiry are not applicable.

The remaining blocker is the original's complete free-block ownership and the menu's decision after delivered input. A future separately authorized run should first validate an exhaustive bounded fragmentation discriminator and cursor-warp delivery control, then choose one original observation. These findings do not authorize an additional original in this task. No shared source was changed; the committed deliverable is this finding, with reproducible run-owned controls.