GetFullPathName directory separator compatibility

Historical source last updated, 2026-10-19. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

The original Army directory-validation-v2 observation is sealed separately in scratch/runs/20261019-army-men-directory-validation-v2, manifest e9dfb850f15bf3dfbeb28832022216587756ac48c6952c7800acb486c0f66a76. It captures a GetFullPathNameA input ending Campaigns/, followed by the game's FindFirstFileA pattern campaigns*.* without a separator, failed enumeration, then a leading-NUL cached directory passed to native _chdir. The original SetAt instruction trace had zero hits; exact dynamic store attribution remains unavailable. No claim that the earlier AVI transition proves this campaign path.

Contract and scope

Microsoft GetFullPathNameA documentation specifies successful length excluding NUL, required capacity including NUL, and a zero file-part pointer for a directory result. It does not require the path to exist. The shared host import serves A and W in TCHAR units.

Wine's Windows path conformance tests, test_ValidPathA lines129–137, explicitly compare a trailing-backslash input's full result including that separator and require NULL filePart. This supplies the trailing-separator evidence not explicit in Microsoft's prose. No original application binary or alternative assets are used in the new tests.

The fix changes API display semantics only: normalize slash spelling before resolving an API path, preserve an explicit final separator, and represent a drive root with its backslash. Canonical VFS lookup still uses the unchanged _resolvePath/_normPath functions; existing case-folding policy is retained. On short buffers, return required TCHAR capacity before writing output or filePart. On successful directory-form output, set filePart to zero; for a file result retain the pointer into the output, scaled for ANSI/UTF-16. Empty inputs retain their existing failure behavior. This is not a complete Windows path normalization overhaul (case retention, device namespace and per-drive CWD limitations remain outside the scoped change).

Validation and pending runtime

node test/test-full-path-name.js passes pure-JavaScript production host imports using16KiB ArrayBuffer memory, no WebAssembly instance: A/W explicit separators, mixed separators, roots, parent traversal, relative/nonexistent paths, sufficient and exact-short/zero capacity, output sentinels, optional and directory/file file-part pointers, and composing Campaigns/ with . for enumeration. node test/test-vfs.js:46 passed,0 failed. Syntax check passes. The sealed old v4 VirtualFS negative control returns C:\work\campaigns without its slash, demonstrating the regression that the new test detects. The first test draft compared a Uint8Array to Buffer with strict equality; that test-only representation mismatch was corrected before all checks passed.

No build, native guest, browser or original Army retry has run on this candidate. Root must review/integrate the disjoint filesystem hunks and qualify an immutable consumer before any claim that this fixes BootCamp gameplay. Original files and historical bundles remain untouched.

Paired original consumer ready (not run)

Immutable /home/user/army-men-local-stage-v7-20261019 uses historical7c943 source/runner903d and module7fd, plus this scoped filesystem JS candidate afeabfd0e06f70b03ab1c8627abacaf76cb1dd66. Maincdc48 filesystem base and v6 filesystem were byte-identical SHAca27bf7e, so the candidate is exactly these reviewed hunks on the prior consumer. No src/ABI or WASM change is required; this is explicitly a paired historical baseline, not currentmain's module. The candidate filesystem SHA is 04f5fd5fa989f416826eb19fe38227bf26a9433d11c921f11bc5f95496dc5fbf.

All1,600 staged entries rehashed,1,597 unchanged hardlinks. Added allocated file bytes434,176 including manifest, excluding directory metadata. Priorv6 also rehashed unchanged. Manifest: b22f8986ce3428e776b34736cb4fe217b373f15511385e090986fc09c6b447a8. Controller: e43ad070c53ac17dd9ca0ca9dc0d76890e03b42847d5ad5b8a5328408be170d1. Unchanged trace plan: 84bb379481ed5ebc5721dd83245ccbfc34e885a9bf6ba911a8eb985e73b4f2ed.

Paired v7v2 result: separator fix validated, game still blocked

Nevertheless snapshot202350 reports the game's Error dialog and the raw log still says ERROR: .AMM does not exist. The framebuffer PNG remains the menu; no gameplay is shown. This trace does not include campaign fopen or chdir return, so successful campaign opening is not proven. Normal HTTP quit exited0 with no guard reason, minfree2,989,023,232, all1,600pins unchanged and freshly absent groups. Prior licence acceptance remains true; Finish unobserved. Evidence16artifacts447,430bytes, manifest ad4488adcf92a901d5eca7b3aa0719d77b21a27a6d2a5f69ef4db99622d2a732 in scratch/runs/20261019-army-men-fullpath-observe-v2. No retry.

Separate source-only reproduction points to a second defect: with cwdc:\work
and an existing c:\work\campaigns directory, _resolvePath('C:/work/campaigns/') returns c:\work\work\campaigns and directory lookup fails, whereas the equivalent backslash absolute path resolves correctly. Drive-relative detection currently runs before slash normalization. This does not retrospectively prove the untraced chdir return; it supplies a generic canonical-resolution regression for a separate source fix and subsequent observation.

Separate canonical slash-classification fix (source only)

Root authorized _resolvePath ownership after the paired separator result. Normalize slash spelling before the existing drive-relative/root/UNC decision, not after it in _normPath. This one-line behavior change keeps C:/dir absolute and /dir rooted, while C:dir remains drive-relative. The new regression first failed on the unmodified function with c:\work\work\campaigns, then passes against the fix. It verifies generic mixed slashes, same/different drive-relative paths, roots, UNC spelling equivalence, chdir to a real mounted directory, relative file opening there, and failure for genuinely missing directories/files. The existing canonical UNC representation is retained, not claimed fully corrected.

Validation: new slash test PASS; A/W API tests PASS; existing VFS46/0 and adopt PASS. Requested lazy suite:49PASS/2FAIL identically on candidate and pre-change HEAD. Both failures construct a ThreadManager prototype without its threads map, then hit threads.values; they are unrelated to path resolution. Baseline was loaded from git HEAD into a test-only Node module cache without modifying production files. Raw baseline/candidate logs are retained at scratch/runs/20261019-army-men-slash-source-tests. An initial third lazy failure was sparse checkout missing host.js; adding exact host.js and test/run.js removed that environmental failure. No scheduler fixture was silently patched. This existing suite includes its real Worker/compiled-adapter check, which passed; no original game or normal build was executed. New test's missing-file assertion uses VirtualFS's internal zero failure (not Win32's external INVALID_HANDLE_VALUE).

The first GetFullPathName fix remains a separate coherent commit; this second fix has no consumer runtime qualification yet and must not be labeled a BootCamp success.

Paired v8 ready; runtime grant still required

Immutable /home/user/army-men-local-stage-v8-20261019 retains historical7c943 runner903d/WASM7fd and original433state/nativeDLLs, with candidate28a2e9035 filesystemSHAba7e94aec476e403c6db8802fd2da724a06eab45b16eee8eb7d0be223b3d287d. Manifestce42e21b5c95fc55ccbc7c9820ac3c28de6f817f9e826f2d0486b19dc354c82c; controllerf1ccf470639ef3bf909f1b42295f0eccdec7760b2da9a10cebd05e55b1c2ed12. All1,600pins checked,1,596 unchanged hardlinks,438,272 newly allocated file bytes including manifest (directory metadata excluded). V7 preserved unchanged. Staged A/W and slash chdir/relative-open tests pass, source95/statics39/native5 preflights pass, and actual trace-parser/host-wrapper synthetic tests pass.

The plan adds selected CreateFileA/ReadFile plus fs_set_current_directory host returns and existing --trace-fs to the previous directory observation. This captures actual resolved file paths/handles/read results and directory success, not merely numeric native-CRT pointers. Existing4MiB log cap stops excess output; no cropped-log success claim. No debugger breakpoint, no automatic menu input. All60/75/90 deadlines,28MiB output,512MiB transient admission and hardfloor remain. Source-only receipt scratch/runs/20261019-army-men-slash-consumer-preflight/ready.json.

V8 original BootCamp scene reached

Actual log10877 changes to the canonical Campaigns directory; host return1. BootCamp.cpn opens as70000028, then yields0x5ce(1486)bytes and EOF. BootCamp.AMM opens as7000002d, with subsequent actual map reads. The selected read trace reports copied bytes, not a separately logged final ReadFile error; downstream original scene progression corroborates successful campaign/map processing. Normal quit exited0, no guards, minfree2,274,381,824, all1,600pins unchanged, groups freshly absent. Historical7c943/7fd plus scoped JS fixes remains the accurate runtime identity; this was not a currentmain WASM run. Prior actual licence acceptance true, installer Finish unobserved.

Evidence scratch/runs/20261019-army-men-slash-observe-v1:17artifacts, 4,343,609bytes before manifest, rawlog3,103,579bytes uncropped; manifest 6e5aa56b55f605568df29532c576f98efbe562270172f87ba91a9ef192349b9d. No retry. Next ordinary movement observation should remove broad filesystem and ReadFile tracing to avoid audio streaming log volume.

Next movement proof, source-only plan

The installed original readme.txt, July1998, explicitly documents Numpad8 for forward movement, Numpad2 backward and Numpad4/6 turning. Space/left-click fires, Enter toggles map-scroll mode, and Escape opens options; those are not movement substitutes. After normal observed menu/BootCamp entry, retain before/after frames around a bounded Numpad8 keydown/keyup (VK104), using ordinary control commands, then inspect Sarge/terrain/minimap change. Stop on failure; no memory writes, timed game workaround or undocumented sign-dismiss input. If the welcome sign prevents a defensible movement comparison, report that limitation instead of claiming success. No next consumer or runtime has been launched yet.

V9 movement consumer ready, no runtime grant yet

Immutable /home/user/army-men-local-stage-v9-20261019 retains V8 runtime, originals and historical module unchanged. Only controller/plan/identity differ: 1597 hardlinks, 282624 added allocated bytes excluding directory metadata. Manifest 738fd64e33e4f5bdb6a7d461a74941ebab59ae152d5fdf4b03bbe9a20f507525; controller 8da0bab81381fbca3a2c3a0f901969e5781e19a6823fec6d8bcf73a7688561af; plan 1f8d1f977815d71a98e5366f1b6112899f7b54cd5a80c3897d387a72ac52f78f. Source and native dependency preflight passed without guest execution; receipts in scratch/runs/20261019-army-men-movement-preflight. Pure JS mocked movement ordering/release tests pass, including failed down acknowledgement and wait. Only --trace-input remains; broad file/audio/API traces are removed. All 60/75/90s, 4MiB log, 28MiB output, 512MiB transient and disk guards remain.

V9 movement-v1 was granted, but the exact controller refused initial fresh disk admission at line28, exit1, before creating its evidence directory or spawning any child. No movement/menu input occurred and no retry was made. The guard remains free>2847932416. The assertion does not record its exact failed sample; scratch/runs/20261019-army-men-movement-v1/admission-failure.json records a separate subsequent statfs observation, not the failed instant. This is an admission refusal, not an original-game compatibility result.

V9 movement-v2: control response, rendering qualification remains partial

Normal quit0, no guards, minfree2260598784, all1600input pins unchanged, survivors[] and both PIDs freshly absent. Rawlog499913bytes (reduced trace); 20artifacts2483001bytes before manifest, retained at scratch/runs/20261019-army-men-movement-v2, manifest 735697be9e720943ab28f923158b2a494e460c9c6cec28f94bb573f159321113. V1 refusal remains separate. Historical runtime and Finish-unobserved limits remain unchanged. No additional input or retry was sent.

Next source-only plan: another separately granted ordinary scene observation can use documented Numpad4/6 turning and Numpad8 forward to leave sign proximity, with explicit releases and a frame after each bounded input. Readme and the observed obstacle sign agree on those controls. Do not use arbitrary dismiss keys. Determine whether stripes persist with the sign absent; if it remains obscured or corrupted, preserve that result rather than declaring gameplay.

V10 remote original: sign clears, terrain rendering failure persists

Granted command cmd_fnsqfnms0l2a, controller18694/guest19384 on bx_rqnkhrnr. Reviewed menu then normal BootCamp320,180 at2268; reviewed welcome scene before Numpad6 turn. Key102 down2996/up3021, frame3025: welcome overlay disappears, Sarge and HUD are visible, but severe diagonal striping spans the terrain. Root independently reviewed the frame. Stopped at this meaningful failure; forward was NOT sent. This establishes a visible control response and persistent rendering defect outside the sign, not clean gameplay qualification.

Normal quit0, no guards, rawlog449547bytes, minfree49145659392, all1600pins unchanged, survivors[], fresh ps bothPIDs absent (cmd_fzuwo5rhi0ag). Evidence recovered with documented download-url+GET API, archive1483489bytes SHA256 b0439c2539e7eb2ce9f0cf42d7b5c36f2b7ece72150f8971975295d7318eea2c, verified before extraction locally to scratch/runs/20261019-army-men-sign-exit-remote-v1. Historical runtime and Finish-unobserved qualifications retained. Runtime released; no retry. Next source-only work must identify the actual rendering/API path from existing logs/source before testing any overlap-scroll/pitch hypothesis; screenshot appearance alone does not justify a renderer patch.

Rendering attribution and bounded next trace, source only

The original PE imports DDRAW.dll!DirectDrawCreate and GDI32!StretchBlt. V10's input-only log contains no surface/Blt/pitch records, so it cannot identify which operation produced the corruption. Inspect the historical V10 source, not the newer main Surface implementation: DirectDraw Blt starts at src/09a8-handlers-directx.wat4999, BltFast5764, Lock6575. BltFast's unkeyed path copies rows in increasing order with $memcpy; the helper uses WASM memory.copy (overlap-safe within one row). Cross-row overlap remains a hypothesis requiring actual same-surface source/destination rectangles. The keyed helper preserves a forward scalar loop for overlapping rows. StretchBlt uses GDI surface descriptors and its separate blit path (09a4 line3003). None of these source observations yet connects the original turn to a defect.

Historical runner --trace-dx supplies decoded rectangles/surface metadata, but it also enables host Lock/Unlock/Present tracing from startup. The trace-api pause control does NOT gate this category; only a static batch window gates it. Turning that broad flag on would not honor a live reviewed turn boundary. The next first discriminator therefore uses only selected --trace-api-paused entries for Blt/BltFast/Lock/Unlock/Flip, StretchBlt/BitBlt, and unresolved . COM dispatch normally resolves ordinal IDs to names before filtering. The helper enables API tracing, sends documented250ms Numpad6, releases, pauses tracing, then captures after-frame/snapshot. It attempts pause even after enable/input/release failure and ordinary quit on failure. No extra movement, host trace, surface export or guest patch.

API memory specs capture receiver/source wrapper bytes and RECTs at entry. Important limits: Lock's entry DDSURFACEDESC is not its returned pitch; the historical table has nargs5/no typed-out descriptors for these COM methods, so BltFast's sixth flags argument is not promised by the ordinary formatter. This first trace discriminates actual API path and same-surface rectangle relationship; it must not claim full DX metadata or correct pitch. A native body bypassing logged entry points remains unobserved. Raw startup leakage and all raw output are preserved, not silently filtered. Same4MiB log/28MiB output and60/75/90 guards stop any unexpected volume; expected call count is unknown before measurement, and trace overhead may alter progress rate.

V11 staged source-only beside V10, both locally and on bx_rqnkhrnr at /home/user/army-men-local-stage-v11-20261019. Manifest 16b731e683d2033c597de880fa8afeb38425705003a61a37d6133bd30713d8ac; controller ae395ad1cc656e36a39d149b5e9f137d3f75bee87f60a2e700c82e20a595d2ea; helper c9c644e2da1721bfaeea2035db245208e20e848cc8daa62981016392c8e9fa21. All1600hashes match,1597hardlinks and282624allocated new bytes excluding directory metadata; V10 pins verified before/after. Only three source files were uploaded via independent Files API (archive4898bytes, SHA16277a2a2e3cca36d440b17b6b3ef9ef2f39e7ff149025d35e89f6687b581d61). Remote staging cmd_nfdbaitmy4ix and verification cmd_prqdp5mxwxbc passed; free48769445888, output parent present, unique output absent and port8143 not listening at02:48:24UTC. Recheck port/admission at allocation; Dungeon owns next runtime slot. No guest/native preflight/runtime was launched.

V11 terminal startup log-cap, no turn observation

Granted remote cmd_kx6fdilkaxly/controller26947/guest27636 passed1600pins, free48743530496 and port8143 clear. It hit the unchanged4MiB log cap after 4.8seconds, before first screenshot request or any menu/turn/helper input. SIGKILL/log-cap, received4205278bytes; retained raw4194296bytes, with the over-cap tail intentionally not retained by the existing guard. No trace-api toggle markers. Startup trace contains5324Blt,688Lock,687Unlock,3Flip entries; these do not establish the turn's rendering path. Pins unchanged, survivors[], minfree48739008512. Fresh ps/controller+child and port8143 clear confirmed in cmd_xmtr20m9gc87. No retry. First later screenshot request correctly failed ECONNREFUSED and download404; no image existed.

All13artifacts4488519bytes recovered through documented Files API to scratch/runs/20261019-army-men-render-trace-remote-v1, sealed manifest 1d98026eb9f5d7e96e6e10d1e67426060c559de359c134065529c5d04098773a. Archive152544bytes SHA4deef68d28a65b90af85b1882d59ff70a3817eb4207495f6a55b4a6195aa8e00. Exact tracing defect found in both historical/current runner: initial apiLogOff predicate ends !TRACE_API_MEM.length, so memory specs defeat TRACE_API_PAUSED and leave the WAT logger on. COM dispatch itself correctly checks api_log_on and resolves ordinal names. The prior parser/toggle mocks did not exercise this initial predicate interaction; new regression must. No renderer change or original-game result follows from this failed trace.

Scoped runner correction changes only the memory-spec term of the initial predicate to (!TRACE_API_MEM.length || (TRACE_API && TRACE_API_PAUSED)). Memory dumps paired with a paused API trace no longer force dispatch logging on. Memory-only tracing without TRACE_API retains its existing behavior even if the unused paused flag is present; trace-api control still rejects that unsupported combination. Counts/critical/input-dispatch/ESP/audit/breakpoint consumers continue to force logging independently. The correction does not promise pause suppression when such independent consumers are enabled.

tools/test-army-men-trace-memory-pause.js extracts the production predicate, COM marker resolver and entry memory logger, and actual trace-api toggle. Synthetic dispatch emits the real COM marker/placeholder only when the set_api_log stand-in is enabled. It verifies startup silence, actual resolved COM memory bytes after on, silence after off, memory-only behavior and each independent consumer. Historical V11 fails the initial-silence assertion; candidate passes. Syntax/diff checks pass. This is pure JS source conformance, not a native COM execution test; original retry remains separately gated.

V12 original turn reached; trace volume still exceeds cap

cmd_jaldjha5xa6h/controller30161/guest30850 reached reviewed menu/BootCamp, click1039. Input receipt confirms traceON1651, key102down1653/up1661, traceOFF1664, afterPNG1667. Severe diagonal corruption remains. It then terminated SIGKILL/log-cap while draining buffered output, before attempted normal quit (which returned ECONNREFUSED). Received4235060bytes, retained raw4167893bytes; minfree48561946624,1600pins unchanged,survivors[]. Fresh cmd_tixem3k0fmpk confirms bothPIDs and8143 absent. No retry.

Initial API pause works in this original execution: no API entries before ON. However9638unrelated [FS]audio-read lines already used~1.89MB beforehand. Retained trace then contains1969BltFast/1599Lock/1599Unlock/123Flip calls, but ends BEFORE the keydown marker. Separate control acknowledgements and afterPNG exist; their corresponding trace tail was not retained. There is no same-wrapper BltFast in this retained prefix; that fact CANNOT rule out a later turn-related same-surface copy. No StretchBlt or Blt in the prefix. Common full scene copy is8221080->8221078, RECT0,0,640,386, callerret4064c8; Lockret420382/Unlockret4203c5 operates on8221078. Lock helper420360 stores returned lpSurface at5cc5ec and pitch at5cca28; entry dumps cannot prove them.

Static original caller406460 invokes407440 to update the scrolling background. At4074f1..407503 it pushes the SAME global4744d8 surface for source/destination, RECTonstack, destinationECX/EDX, and flags0x10, then calls vtable+1c(BltFast). This is a concrete candidate call site, not proof it ran in the retained turn trace. A bounded existing trace-at407503 capturing stack/RECT/registers with small hit limit would distinguish overlap directions without broad API flood. No renderer patch is justified before actual arguments and semantics review.

Evidence recovered via Files API archive1118995bytes, SHA575f093fb4def00dd9cf00bc8a9133b30948966fb6f07767aa0ae44fa9685c22, at scratch/runs/20261019-army-men-render-trace-remote-v2. All prior partial gameplay, original acceptance and Finish-unobserved qualifications persist.

V13 prepares only --trace-at=0x407503 --trace-at-limit=8, memory esp:24,*esp:8,*esp+12:8,*esp+16:16,0x5d91c4:32, plus input tracing. At this pre-call boundary stack DWORDs are receiver,x,y,source,RECTpointer,flags; the two wrappers and pointed RECT are captured using the actual resolver. No API/hostDX tracing or forward input. The first eight hits may occur before turn; correlate batch with input receipt, and treat exhausted/no hits as unavailable for the turn rather than infer execution. Test executes actual pointer parser and ordinary250ms turn protocol with synthetic data; PASS.

V13 granted cmd_vh4ax9tj3ztx/controller34284/guest35000 passed immediate1600pins, free48321064960 and port check. Reviewed menu/BootCamp, click409, ordinary key102down789/up801, after804 still severe diagonal striping. ZERO407503 trace hits, not eight exhausted hits. Actual copy arguments remain unavailable; no execution/nonexecution or overlap conclusion is justified by the missing capture. Normal quit0, no guards, raw143670bytes, minfree48318521344, all1600pins unchanged,survivors[]. Fresh cmd_vbijnlxl5t2n confirms controller, child and8143 absent. No retry or renderer patch. Archive1454098bytes SHA075eed73454f01d15d9274c103723e17adc7f64c202f50e9794fd105139ae6e0 recovered/verified via Files API to scratch/runs/20261019-army-men-scroll-boundary-remote-v1; all prior originals and evidence retained. Next source investigation must explain whether this instruction boundary is observable with the existing debugger before another game allocation; repeated no-hit runs would not resolve the renderer cause.

Source observability diagnosis: historical13-exports.set_bp only stores the address and recomputes debugger chain/tier guards. $run compares EIP at block dispatch; no bp_addr check splits decoder instruction streams, and set_bp does not invalidate/split cached blocks. Block-executor guards cover interior block entries, not arbitrary instructions inside one block. 407503 is inside the branch-target block4074f1 (jne4074e2 targets4074f1), so this explains why the selected instruction is not a reliable trace point. Original PE preferredbase400000/entry46fee0 and pinned disassembly agree; this is not evidence of a rebased DLL address mistake.

Synthetic PE/x86 only: compare a breakpoint inside straight-line instructions with a real branch entry, then test resume/rearm and cached branch-entry hit. Syntax passed; native execution NOT yet granted/run. Nearest proposed original boundary is 4074f1: ECX/EDX already contain destination coordinates, ESP points to source RECT, global4744d8 supplies both receiver/source, following pushes supply flags0x10. Do not run that original observation until the synthetic control demonstrates the chosen boundary method. No forced split/core patch proposed.