Aranna selected-map loader investigation

04:33 successor: filename case repair published

Commit 64823a6ad8bc73ada23f39a1f61573172d10837c fixes a demonstrated GetFullPathNameA/W defect: display paths lowercased caller-supplied filenames. Lookup keys remain case-insensitive. Both API variants now retain the filename spelling, including nonexistent files, dot traversal and drive-relative paths. The focused regression failed before the repair and passes afterward; all 46 VFS tests and the mandatory locked build passed. Evidence for this successor is under scratch/runs/20261107-aranna-map-load/successor-0433/; the earlier 01:39 attempt below and its evidence are retained.

The acquisition-before-arm predecessor's ordered observations establish a more precise failure than its dialog alone:

This identifies a concrete emulator discrepancy on the executed path. The filename bytes at the final substring call were not directly captured; that link follows the static data flow and demonstrated API behavior. It does not establish that this is the only gameplay blocker. Microsoft documents that GetFullPathName does not convert the specified filename. No guest, map, resource, save or configuration bytes were changed. There were zero read mismatches in all 28,643 retained broker completions; the earlier optional startup misses are not assigned as the map cause.

The fresh-box attempt stopped before guest launch. Box bx_dbhdzfqg was created at 04:38:49Z with a persisted UUID Idempotency-Key and 40-minute TTL (original expiry 05:18:49Z). Its prerequisite check verified all 1,906 staged file size/SHA-256 pins, then failed at require.resolve('puppeteer') with MODULE_NOT_FOUND. No original-attempt marker, browser, guest process, profile or app port was created. No install/retry, alternative driver, original replay or old-box resume followed that concrete blocker. Therefore the repair's effect on the original map load, world rendering and visible movement remains unverified. Next work must prepare the missing browser-driver dependency before a separately authorized original on the published repair.

The exact preflight failure and remote cleanup were recovered before the explicitly delegated normal shutdown request at 04:40:20Z. Remote cleanup rechecked all pins and found no owned processes or profiles; listed listeners belonged to the base box. No local runtime was launched. Source/build ownership was released after publication. Final box state and local cleanup are retained in this successor's result and lifecycle JSON. No source/media archive, deployment, nested worker or agent was created. The observations and generic repair tests do not constitute a new original-gameplay pass.

Earlier 01:39 attempt (retained)

Evidence lives in scratch/runs/20261107-aranna-map-load/. This follows the complete-transport predecessor. The one new original ended at Single Player before Start New Game: an unnecessary canvas-acquisition command while pointer lock remained active was rejected by the input guard. This was the worker's input-selection error. No map-load cause, production repair, loaded world or player movement is established.

Original result and exact blocker

The original ran from 2026-11-07T01:39:02Z to 01:44:31.201Z, within its 40-minute limit. I personally reviewed EULA, intro, Main Menu and Single Player. A fresh native proof and complete ordinary balanced click selected Single Player at native (399,155). The final reviewed scene is original/response-12.png.

At 01:44:30.592Z, command 13 requested canvasclick:0.5,0.5 although its reviewed ticket already said mode=relative, lock=screen. The unchanged input-adapter.cjs:61 requires a canvas click to be an initial browser click or lock acquisition. It asserted “explicit reviewed browser click only” before any mouse operation. original/action-13.json records rejected-terminal, effectsAttempted=0, and no button receipt. This was a terminal assertion, not the recoverable stale-input refusal path. No new ID, replay, resumed original or second attempt followed. The driver exited 1; its browser exited normally 0.

The final file-observer coverage was 26,692 broker records, zero drops, zero copied-byte mismatches and no observer error. There were two earlier actual failed opens: C:\DungeonSiege.ini and c:\dir.lqd20, each OPEN_EXISTING, ERROR_FILE_NOT_FOUND (2), INVALID_HANDLE_VALUE. They preceded successful entry to the menus; neither is established as the cause of the predecessor's later map failure. No map-loader decision or post-entry memory-API record occurred because this original never reached that code (original-analysis.json). Native journal maximum count was 1,175, below 8,192; malformed records were absent. All 235 native transactions were complete (original-packets.json). Resource-journal saturation did not occur.

This task's concrete blocker is the terminal input-guard refusal above. The requested first failing map-load API remains unidentified, rather than being inferred from a dialog. No failing-before emulator regression or emulator repair is claimed. A future authorized original would need to follow the already-active relative route with measured movement rather than reacquiring the canvas. The current attempt is terminal and must not be resumed.

All recorded local process trees, HTTP/debug ports and temporary Chrome profiles were independently verified absent in final-cleanup.json. The runtime holder was released, its PID verified absent, and all four runtime/browser/heavy/build locks verified free in runtime-release.json. Runtime/source/build ownership was explicitly released on the messageboard. No boat, deployment, guest/media patch, automated risk flag, FPS/audio claim or backend parity claim occurred.

The predecessor selected ordinary Normal and returned to Choose Hero with “The selected map could not be loaded.” Its journal contains no file completion or allocation-result observations at that failure. There is no post-Normal HTTP failure or heap-OOM notification. Those absences do not prove successful file access or allocation. The input coverage was complete; this investigation does not reopen the resolved journal-saturation work.

Offline disassembly locates two callers of the error string at 0x77d6f0. The 0x482155 call to 0x5babc9 tests AL at 0x48215a. The other caller tests 0x5babc9 at 0x4bb378 and subsequently tests 0x51f06a at 0x4bb387. Function 0x51f06a checks a registry entry for Expansion.dsmap and three header words: offset 0x3c = 0x47504721, 0x50 = 0x8e8d717a, 0x54 = 0x81665c2e. The canonical map matches. This rules out a mismatch in those canonical bytes, but does not establish which branch ran or what the guest actually read.

The filesystem implementation is lib/filesystem.js; the suggested lib/host-files.js does not exist. ReadFile count/error behavior, signed SetFilePointer handling and WAT VirtualAlloc were compared with the ReadFile, SetFilePointer and VirtualAlloc contracts. A possible semantic difference without a demonstrated failing call is insufficient grounds for changing the emulator.

Run-local passive WAT diagnostics add bounded map-decision snapshots and selected file/memory API results. A host observer copies actual ordered broker open/read/seek completions, comparing returned resource bytes with their VFS source without changing return values, exceptions, position or guest bytes. Production Worker URLs and broker ownership are preserved.

Changed controls passed:

Observer module SHA-256:

Module SHA-256
standard 51a367dd9d6d7f06ecd27f08abd948b85cd83dd12fdf69172044232cbc56d930
compat ab2de6f62a52adc381732537447435c32ed5264dd40c7eb3fdc0104312ca3ea2

Canonical extracted-media identity is preserved, not native-install provenance. The unchanged before-load mounting is inherited from the original-local run. No executable, resource, save or configuration is modified to force progress. The 28 referenced media hashes are verified in retained-audit.json:

File SHA-256
dsloademo.exe e3923f18d3ec9bf52ffe556821dbaa2078e985087c7bf4b171a36b9883b851cf
DSLOA.exe fced0961231248c09c10d86ff51935569db5f83ed3b2ba2739f0501dbc318589
Expansion.dsmap c397bede7fc61ad8650bc835caa6a3fdadd3a3d1c8e3199bc8b1772c2db6dcdc
Expansion.dsres 3ed0688349f33b2dd7b9458c18cfbbc00490353c5888a93cd02c10da7f3c3b10
Logic.dsres 90b26f179364a4d25f7053e8aa41932d51fcd0e563e37820894c3dbfdcf9d0e5

Git SHA plus relevant dirty binary diffs and owned new helper patches are the source archive (source-identity.json, plus source-untracked.json and its patches for relevant untracked runtime sources); no source-tree or media archive is made. The current production modules include the independently published Lords2 PIT repair. The original served commit was cd916e6ef1840055bd97f18f728d7f993fe31515; every pre-original source/helper pin remained identical at the post-original check, with only the new offline analysis helper added. original-http-identity.json verifies the served source/media, and source-post-original.json records the comparison. No production code changed, so this documentation-only closeout does not claim a new locked build.