fix(build): preserve quoted third-party driver paths #1021

Open
NitnekLeBorgne wants to merge 1 commit from NitnekLeBorgne/Longshot-contributions:fix/third-party-driver-paths-20261116 into main
Contributor

Type: fix · Risk: low · Tested on: Linux/native Windows game-free Python fixtures; native Windows Clang 23.1.5 and Linux Clang 18.1.3 driver transcripts (GPU not used) · Needs: full stamp/preprocessing and object-reuse validation

Why

The release's precompiled third-party objects are matched using compiler metadata that should not depend on where the build tools are installed. Splitting a quoted Clang argument at every space leaves bare directory fragments in that metadata. With C:/Program Files/Longshot Build Tools, unchanged main retains Build fragments, so otherwise identical installations do not match.

This is a reproduced build-metadata defect, not a measured runtime or player incident. The parent's public corpus and complete file-list ownership checks admitted this source and new test; no matching issue is claimed.

How to reproduce

Seen on: Linux using explicitly synthetic Clang -### replies; no GPU or compiler involved.
Settings and switches: no new switches; same cc1 options in each fixture.
Steps: run python -I -B tests/test_third_party_driver_paths.py with the new test on main d9486a0846e90905e2ee0bdc4a0d6df96c547c87 and on this candidate.
Expected: C:/Tools and C:/Program Files/Longshot Build Tools yield identical driver-stamp bytes.
Actual: main fails byte equality; the spaced installation leaves directory fragments. The initial focused RED was one assertion failure, no errors.

What changed

Quoted compiler arguments are now kept together before the existing path filter runs. Identical options produce path-independent metadata even when a tool installation contains spaces.

  • nx2/third_party_objects.py, _driver(): tokenize whole double-quoted Clang arguments before the existing path filter, retaining Windows backslashes rather than applying POSIX shell unescaping.
  • New tests/test_third_party_driver_paths.py: seven game-free tests cover the regression, Linux/Unicode and Windows slash forms, escaped quotes, unchanged no-space bytes, distinct CPU/target/feature/default values, quoted semantic values, and existing failed/missing-cc1 behavior.
  • No other production files or existing tests changed. In particular, tests/test_third_party_objects.py remains unchanged (neighbor PR #966 owns it).

Measurements

The reproduced byte-equality regression is fixed. No game or build-time performance was measured; this changes only compiler metadata parsing. No FPS, player startup, compilation, linking or object-cache hit-rate claim. Native driver-only invocations are described below.

Flags and settings

None.

Testing

Additional native Windows checks: CPython 3.14.7 on the exact head above: python -I -B tests/test_third_party_driver_paths.py -v passes seven new methods and python -I -B tests/test_third_party_objects.py -v passes all fifteen unchanged methods; both native exits 0. Those fixture compiler replies remain synthetic. A separate, real Windows Clang driver differential is described below.

Tested: candidate 2370dd31fdc1aa330dff666108cdadde3159984b, tree 880e51bb55af63f1683c62c1e768f7c328c453c8, direct parent d9486a0846e90905e2ee0bdc4a0d6df96c547c87.

  • Focused regression: RED before production edits, GREEN after the minimal patch.
  • Seven new tests and all 15 unchanged third-party-object tests pass. The old tests emit pre-existing ResourceWarnings in nx2/build.py when run verbosely; no warnings were suppressed.
  • Exact committed source: normal python -I -B tests/run_all.py --jobs 3, no selection or skip overrides, Linux, isolated HOME/XDG/TMP and baseline-identical PATH. Result: 538 passed, 22 failed, 92 skipped, 652 checks, exit 1, 230.272 seconds wall time. Four discovered C++ checks pass.
  • Frozen baseline reused without rerunning: 537 passed, 22 failed, 92 skipped, 651 checks, exit 1. Every existing check status, failure case ID and skip reason matches; only the new passing test was added.
  • Strict failure-signature comparator remains false because two missing-Tooling diagnostics embed checkout-specific sibling paths (test_bug_report.py, test_hardware_report.py). Their text is identical after substituting only those exact paths; a separate no-regression assessment passes. Original strict outcome and diff evidence retained; this is not CI-green.
  • Source bytes and POSIX/Git executable modes unchanged across qualification, full parent tree preserved outside the two allowed files, process group empty after exit, reserves monitored above 3 GiB, shared heavy flock used.
  • AGENTS.md was never materialized; two unchanged upstream SwitchesRemoved checks additionally passed with the public blob supplied only in memory. This is normal native-runner qualification of a source-only checkout, not exact-tree CI.

Existing baseline failures retained:

  • tests/test_appimage_player_build.py
  • tests/test_bake_portability.py
  • tests/test_bug_report.py
  • tests/test_crash_dialog.py
  • tests/test_crt_options.py
  • tests/test_env_on.py
  • tests/test_guest_imported.py
  • tests/test_hardware_report.py
  • tests/test_lift_cleanup_error.py
  • tests/test_link_skip.py
  • tests/test_linux_update.py
  • tests/test_make_game_cache.py
  • tests/test_missing_nup.py
  • tests/test_mods_size_table.py
  • tests/test_options_budget_once.py
  • tests/test_pending_guest_import.py
  • tests/test_pipeline_tour.py
  • tests/test_readback_late.py
  • tests/test_run_all_changed.py
  • tests/test_vk_occlusion_queries.py
  • tests/test_vk_pgo.py
  • tests/test_walk_key.py

Earlier real native Windows resource-argument differential (2026-11-16; not physical relocation): Exact published head 2370dd31fdc1aa330dff666108cdadde3159984b, unchanged nx2/third_party_objects.py SHA-256 105397297ee1d8a6b4dbab981febb77f5a24d2ba2d56087e8ae2d05ad369e4dd, normally imported after static confirmation that its top-level closure is stdlib-only and does not build/download. Explicit existing CPython 3.14.7 and Clang 23.1.5 were used; compiler SHA-256 d05fee2c3f79b15d998d3c3413c76aabd416ec789edff89c8f698a8be44a3970.

  • The unchanged real _driver() invoked Clang -### with --target=x86_64-w64-windows-gnu -march=x86-64-v3 -nostdinc, comparing -resource-dir D:/PublicTools/resource with -resource-dir "D:/Public Tools épreuve/resource". The latter contains one three-word Unicode directory component. This is an actual resource-directory argument axis, not two relocated compiler installations; the directories need not contain headers for this driver-only check.
  • Real C and C++ cc1 transcripts give unequal metadata on the frozen direct parent and equal metadata on this candidate. No-space C/C++ bytes remain identical to the parent's. Separate real CPU (x86-64-v2), feature (-mno-avx2) and target-triple mutations remain distinct, and the actual quoted nonpath macro value PUBLIC_LABEL=hello wide world is preserved.
  • The final three-word-path run contains 16 real driver calls plus one real --version call, all native exit 0; harness exit 0. An earlier four-word Unicode path run independently passed the same controls and is retained separately. Complete raw native stdout/stderr, argv, return codes, stamp bytes and SHA-256 readbacks are retained, including the literal parent RED differential. Instrumentation records actual compiler output; it supplies no synthetic compiler reply.
  • Production source, original compiler and the full existing private Python runtime file manifest are unchanged before/after; -I -B was used, without child Python invocations. Resource checks sampled before/after native commands retained the prescribed reserves; this is not a continuous-monitoring claim. Subsequent process readback found no matching owned test Python/driver process.
  • This validates only _driver() parsing of actual Clang output. -### does not compile, link, check headers, exercise the complete preprocessing stamp(), run a game or use a GPU. No installation, download, compiler copy, game input or product source edit was needed. Existing Linux full-suite counts, exit 1 and strict-comparator false above remain unchanged; the suite was not rerun.

Earlier real native Linux resource-argument differential (2026-11-16; not physical relocation): The same exact published head and source SHA-256 above were checked against the same frozen direct parent using the existing Ubuntu Clang 18.1.3 (1ubuntu1), compiler SHA-256 8ef402d453d1ba4902e4ee0f0f847f6cfa01400c95aa43c24e97818b9c0e3f45, and native CPython 3.12.3. The existing ELF compiler matched the prior host-659 frozen artifact manifest; nothing was installed, downloaded or copied from the compiler installation. The complete public production module was normally imported after its stdlib-only top-level closure was read; the unchanged real _driver() executed every native driver call.

  • Real Clang -### used --target=x86_64-unknown-linux-gnu -march=x86-64-v3 -nostdinc and two actual -resource-dir arguments under the new evidence root: PublicTools/resource and Public Tools épreuve/resource (one three-word Unicode component). This is the resource-directory argument axis, not a relocated compiler installation, and does not require headers to exist.
  • In both C and C++, the actual parent metadata differs: the spaced path leaks the literal bare word Tools. Candidate metadata is equal across those paths; no-space C/C++ bytes remain byte-identical to the parent's. Separate real CPU (x86-64-v2), feature (-mno-avx2), target (x86_64-w64-windows-gnu) and option (-fno-stack-protector -fno-pic) mutations remain distinct. The actual quoted nonpath macro value PUBLIC_LABEL=hello wide world survives.
  • The Linux run made 18 real driver calls plus one real --version call, all native exit 0; harness exit 0. Two read-only Git blob reads and one ldd dependency discovery also exited 0 (22 subprocesses total). Complete argv, UTC timing, raw stdout/stderr, native return codes, stamp bytes and hashes, and the real parent differential are retained in the local linux-driver evidence packet. The recording adapter delegates to real subprocesses; it supplies no mock compiler response.
  • Production/test source, compiler, native Python executable and every resolved shared-library file were SHA-256-identical before/after. The compiler hash also matches its frozen host-659 manifest. python -I -B was used; child commands had isolated HOME/XDG/TMP, a process-scoped library search path and individual owned process groups with 20-second timeouts. The shared heavy lock was held. Before/after samples for every invocation stayed above 3 GiB disk and MemAvailable; these are samples, not continuous monitoring. All owned child groups were absent after exit, with an independent post-run process readback retained.
  • Windows Clang 23.1.5 and Linux Clang 18.1.3 are separately qualified within each OS. No cross-version or cross-OS equality of normalized bytes is expected or claimed. The earlier Windows 32 driver calls and two version calls are preserved unchanged and were not rerun. No source edit, full-suite rerun or frozen-baseline rerun was needed. -### does not compile, link, validate headers, invoke the complete preprocessing stamp(), run a game or use a GPU.

New physical Linux placement checks (2026-11-16): On the same unchanged candidate and frozen parent, real ELF Clang 18.1.3 was independently copied into NoSpaceClang and Trois Mots épreuve, with 17 resolved libraries and 237 existing resource files per prefix (510 files, 487926162 bytes total). These are distinct physical executable/dependency/resource copies, not resource-argument substitutions; paired runs used identical cwd, environment and options, without a -resource-dir override. Across eight configurations, 32 real driver calls plus 11 other native calls all exited 0. C and C++ path baselines are candidate-equal, parent-RED and no-space byte-unchanged; C-only CPU, feature, target, default-option and quoted-value controls remain distinct. Source and original/copied tools were unchanged. The OS loader and kernel remain shared; copied resource/header existence is not header validation.

New physical Windows placement checks (2026-11-16): Existing LLVM-mingw Clang 23.1.5's real wrapper, clang-23.exe backend, recursive DLL/config closure and resources were independently copied into no-space and three-word Unicode prefixes (924 files, 469882234 bytes total). The 11 actual config files and 10 same-folder includes were kept unchanged. Initial 32 driver calls plus four version/resource calls all exited 0; C and C++ path baselines are candidate-equal, parent-RED and no-space byte-unchanged, while C-only CPU, feature, target and quoted-value controls remain distinct. The original initial PRODUCT_RED is retained: two default-option oracle assertions expected -fno-stack-protector -fno-pic to change metadata, but the raw cc1 arguments were already identical before parsing. A separate additional C-only -fstack-protector-strong control made four real calls and confirmed meaningful raw and normalized differences at both placements, candidate path equality, parent RED and unchanged no-space bytes. It did not recopy tools, replay initial cases or modify original evidence/inputs. Combined Windows totals are 36 driver calls and 40 native wrapper invocations, not a claim about actual process count; the separate three benign Job controls are not product tests.

Scope and retained limits: These new physical-placement checks are separate from the earlier resource-argument checks, not replacements or full-suite reruns. Linux's original global-process audit FAIL remains retained because some /proc entries were unreadable; independent SPEC and QUALITY adopted only the scoped evidence, including absence of all 43 owned child groups. Windows owned-Job and readable own-namespace readbacks were empty; there is no global host process-absence or runtime DLL-load-trace claim. The combined Windows assessment is scoped to these retained receipts, not maintainer approval or a full-suite pass. Neither OS result establishes installation/SDK completeness, full stamp(), transitive preprocessing, header usability, compilation, linking, object reuse or gameplay. Different Clang versions are qualified separately; no cross-OS/version metadata equality is claimed. The existing Linux full-suite exit 1, all 22 failures and strict-comparator false remain unchanged.

Retained harness history: The first Linux placement harness stopped on a lexical ldd-path membership assertion; a scoped continuation reused the same verified physical copies and corrected only that bookkeeping check, without a product edit or recopy. Before Windows compiler execution, review corrected launch containment (suspended creation, owned-Job assignment, then resume) and lossless diagnostic-byte packaging; benign exit/nonzero/descendant-timeout controls exercised that harness separately. The first Windows inspection also stopped on same-folder config includes until the actual finite config closure was reviewed and pinned. The separate four-call control had a failure-path raw-packaging correction reviewed before execution, and an invocation without its required arming flag stopped before any native call. Those review, harness and fail-closed preflight receipts remain retained alongside the initial results; none is erased, counted as a compiler pass or used to change product code.

Not tested: full stamp() preprocessing/header behavior, a native Windows build, shipped third-party-object reuse, or any game. Compiler replies in the original focused tests remain labelled synthetic fixtures; the separate native driver checks above are not mocked.

Must be tested:

  • Windows and Linux, any GPU: by hand, validate complete stamp() preprocessing with usable headers and shipped third-party-object reuse, expecting unchanged inputs/options to select compatible metadata and objects and semantic changes to remain distinguishable. Physical driver placement is now exercised within the scoped checks above; that does not establish these remaining build-level behaviors. No test-farm request has been made.

Risks

What could break: parsing a Clang argument transcript unlike the quoted/escaped form covered here could still prevent reuse. The path filter's existing semantics are unchanged, not broadened to remove semantic options.

Turning it off: revert the commit.

Compatibility: no new flags, defaults, manifest format, runtime code, saves, renderer ABI or shader/pipeline format. Previously contaminated spaced-path metadata changes, intentionally, to the existing path-independent value; no-space metadata is tested byte-for-byte unchanged.

Checklist

Everyone

  • One topic.
  • The title is a Conventional Commit (type(scope): description) of the status line's type.
  • Why gives the reason plainly and links the issues that fit (the public tracker and complete public PR file lists searched; parent scope adopted, no matching issue claimed).
  • Measurements has a verdict, or one line saying why there is nothing to measure.
  • Testing names what was tested, what wasn't, and what must still be tested, where and how.
  • Rebased on current main at frozen intake (d9486a0846e90905e2ee0bdc4a0d6df96c547c87). Public main has since advanced; no rebase to that later tip is claimed. Relevant source closure is checked again before publication, without transferring full-suite results to another tree.
  • python -I -B tests/run_all.py qualified against its exact pinned-main baseline: native exit 1 is preserved, all 22 existing failures are named above, and no new failure or changed skip status was found. This does not assert a wholly passing suite.
  • No game files, dumps, build output or generated files.
  • third_party/ untouched (a new folder with a PROVENANCE.md instead).
  • The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled.

When it applies

  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.
  • A game-free test added for the fix, failing on main and passing here.
  • A check under tests/ that reads changed source updated in this PR, not weakened (new independent file).

Source export scope: unchanged gpu/build-shadertool.bat has the recorded .gitattributes LF-to-CRLF conversion. Candidate production and test files match native Git blobs; source/modes are unchanged across each run. Tooling is absent in the isolated runner context: two corresponding failure messages require only an exact sibling-path substitution; the original strict comparison remains false. This is not hosted-CI or literal exact-tree equivalence.

Windows qualification caveats: Native CPython 3.14.7 used a new independent copy of an existing pinned-dependency runtime, not a venv or a replacement product runner. Each candidate used its own full frozen-head public Git-blob export (2652 tree entries; 2651 files materialized; AGENTS.md kept opaque and never materialized). Production/test SHA-256 and all exported file bytes were verified before/after; Git modes remain metadata, not a claim of POSIX permission semantics on NTFS. Normal public test sources and commands were unchanged. Full stdout/stderr and exact native executable/argv/cwd/exit records are retained. A post-test harness assertion exited 1 because five CPython encoding bytecode cache files appeared only in the new owned runtime copy; independent readback confirmed no existing runtime file changed, the original runtime stayed byte-identical, and all test invocations exited 0. That earlier fixture run did not qualify a real compiler. The separate native driver-only checks above do; no full Windows suite, compilation, game, GUI, hosted CI or publication qualification is claimed. Existing Linux full-suite results above are unchanged and are not transferred to Windows.

**Type:** fix · **Risk:** low · **Tested on:** Linux/native Windows game-free Python fixtures; native Windows Clang 23.1.5 and Linux Clang 18.1.3 driver transcripts (GPU not used) · **Needs:** full stamp/preprocessing and object-reuse validation ## Why The release's precompiled third-party objects are matched using compiler metadata that should not depend on where the build tools are installed. Splitting a quoted Clang argument at every space leaves bare directory fragments in that metadata. With `C:/Program Files/Longshot Build Tools`, unchanged main retains `Build` fragments, so otherwise identical installations do not match. This is a reproduced build-metadata defect, not a measured runtime or player incident. The parent's public corpus and complete file-list ownership checks admitted this source and new test; no matching issue is claimed. ## How to reproduce **Seen on:** Linux using explicitly synthetic Clang `-###` replies; no GPU or compiler involved. **Settings and switches:** no new switches; same cc1 options in each fixture. **Steps:** run `python -I -B tests/test_third_party_driver_paths.py` with the new test on main `d9486a0846e90905e2ee0bdc4a0d6df96c547c87` and on this candidate. **Expected:** `C:/Tools` and `C:/Program Files/Longshot Build Tools` yield identical driver-stamp bytes. **Actual:** main fails byte equality; the spaced installation leaves directory fragments. The initial focused RED was one assertion failure, no errors. ## What changed Quoted compiler arguments are now kept together before the existing path filter runs. Identical options produce path-independent metadata even when a tool installation contains spaces. - `nx2/third_party_objects.py`, `_driver()`: tokenize whole double-quoted Clang arguments before the existing path filter, retaining Windows backslashes rather than applying POSIX shell unescaping. - New `tests/test_third_party_driver_paths.py`: seven game-free tests cover the regression, Linux/Unicode and Windows slash forms, escaped quotes, unchanged no-space bytes, distinct CPU/target/feature/default values, quoted semantic values, and existing failed/missing-cc1 behavior. - No other production files or existing tests changed. In particular, `tests/test_third_party_objects.py` remains unchanged (neighbor PR #966 owns it). ## Measurements The reproduced byte-equality regression is fixed. No game or build-time performance was measured; this changes only compiler metadata parsing. No FPS, player startup, compilation, linking or object-cache hit-rate claim. Native driver-only invocations are described below. ### Flags and settings None. ## Testing **Additional native Windows checks:** CPython 3.14.7 on the exact head above: `python -I -B tests/test_third_party_driver_paths.py -v` passes seven new methods and `python -I -B tests/test_third_party_objects.py -v` passes all fifteen unchanged methods; both native exits 0. Those fixture compiler replies remain synthetic. A separate, real Windows Clang driver differential is described below. **Tested:** candidate `2370dd31fdc1aa330dff666108cdadde3159984b`, tree `880e51bb55af63f1683c62c1e768f7c328c453c8`, direct parent `d9486a0846e90905e2ee0bdc4a0d6df96c547c87`. - Focused regression: RED before production edits, GREEN after the minimal patch. - Seven new tests and all 15 unchanged third-party-object tests pass. The old tests emit pre-existing ResourceWarnings in `nx2/build.py` when run verbosely; no warnings were suppressed. - Exact committed source: normal `python -I -B tests/run_all.py --jobs 3`, no selection or skip overrides, Linux, isolated HOME/XDG/TMP and baseline-identical PATH. Result: **538 passed, 22 failed, 92 skipped**, 652 checks, exit 1, 230.272 seconds wall time. Four discovered C++ checks pass. - Frozen baseline reused without rerunning: **537 passed, 22 failed, 92 skipped**, 651 checks, exit 1. Every existing check status, failure case ID and skip reason matches; only the new passing test was added. - Strict failure-signature comparator remains **false** because two missing-Tooling diagnostics embed checkout-specific sibling paths (`test_bug_report.py`, `test_hardware_report.py`). Their text is identical after substituting only those exact paths; a separate no-regression assessment passes. Original strict outcome and diff evidence retained; this is not CI-green. - Source bytes and POSIX/Git executable modes unchanged across qualification, full parent tree preserved outside the two allowed files, process group empty after exit, reserves monitored above 3 GiB, shared heavy flock used. - `AGENTS.md` was never materialized; two unchanged upstream SwitchesRemoved checks additionally passed with the public blob supplied only in memory. This is normal native-runner qualification of a source-only checkout, not exact-tree CI. Existing baseline failures retained: - `tests/test_appimage_player_build.py` - `tests/test_bake_portability.py` - `tests/test_bug_report.py` - `tests/test_crash_dialog.py` - `tests/test_crt_options.py` - `tests/test_env_on.py` - `tests/test_guest_imported.py` - `tests/test_hardware_report.py` - `tests/test_lift_cleanup_error.py` - `tests/test_link_skip.py` - `tests/test_linux_update.py` - `tests/test_make_game_cache.py` - `tests/test_missing_nup.py` - `tests/test_mods_size_table.py` - `tests/test_options_budget_once.py` - `tests/test_pending_guest_import.py` - `tests/test_pipeline_tour.py` - `tests/test_readback_late.py` - `tests/test_run_all_changed.py` - `tests/test_vk_occlusion_queries.py` - `tests/test_vk_pgo.py` - `tests/test_walk_key.py` **Earlier real native Windows resource-argument differential (2026-11-16; not physical relocation):** Exact published head `2370dd31fdc1aa330dff666108cdadde3159984b`, unchanged `nx2/third_party_objects.py` SHA-256 `105397297ee1d8a6b4dbab981febb77f5a24d2ba2d56087e8ae2d05ad369e4dd`, normally imported after static confirmation that its top-level closure is stdlib-only and does not build/download. Explicit existing CPython 3.14.7 and Clang 23.1.5 were used; compiler SHA-256 `d05fee2c3f79b15d998d3c3413c76aabd416ec789edff89c8f698a8be44a3970`. - The unchanged real `_driver()` invoked Clang `-###` with `--target=x86_64-w64-windows-gnu -march=x86-64-v3 -nostdinc`, comparing `-resource-dir D:/PublicTools/resource` with `-resource-dir "D:/Public Tools épreuve/resource"`. The latter contains one three-word Unicode directory component. This is an actual **resource-directory argument axis**, not two relocated compiler installations; the directories need not contain headers for this driver-only check. - Real C and C++ cc1 transcripts give unequal metadata on the frozen direct parent and equal metadata on this candidate. No-space C/C++ bytes remain identical to the parent's. Separate real CPU (`x86-64-v2`), feature (`-mno-avx2`) and target-triple mutations remain distinct, and the actual quoted nonpath macro value `PUBLIC_LABEL=hello wide world` is preserved. - The final three-word-path run contains 16 real driver calls plus one real `--version` call, all native exit 0; harness exit 0. An earlier four-word Unicode path run independently passed the same controls and is retained separately. Complete raw native stdout/stderr, argv, return codes, stamp bytes and SHA-256 readbacks are retained, including the literal parent RED differential. Instrumentation records actual compiler output; it supplies no synthetic compiler reply. - Production source, original compiler and the full existing private Python runtime file manifest are unchanged before/after; `-I -B` was used, without child Python invocations. Resource checks sampled before/after native commands retained the prescribed reserves; this is not a continuous-monitoring claim. Subsequent process readback found no matching owned test Python/driver process. - This validates only `_driver()` parsing of actual Clang output. `-###` does not compile, link, check headers, exercise the complete preprocessing `stamp()`, run a game or use a GPU. No installation, download, compiler copy, game input or product source edit was needed. Existing Linux full-suite counts, exit 1 and strict-comparator false above remain unchanged; the suite was not rerun. **Earlier real native Linux resource-argument differential (2026-11-16; not physical relocation):** The same exact published head and source SHA-256 above were checked against the same frozen direct parent using the existing Ubuntu Clang 18.1.3 (1ubuntu1), compiler SHA-256 `8ef402d453d1ba4902e4ee0f0f847f6cfa01400c95aa43c24e97818b9c0e3f45`, and native CPython 3.12.3. The existing ELF compiler matched the prior host-659 frozen artifact manifest; nothing was installed, downloaded or copied from the compiler installation. The complete public production module was normally imported after its stdlib-only top-level closure was read; the unchanged real `_driver()` executed every native driver call. - Real Clang `-###` used `--target=x86_64-unknown-linux-gnu -march=x86-64-v3 -nostdinc` and two actual `-resource-dir` arguments under the new evidence root: `PublicTools/resource` and `Public Tools épreuve/resource` (one three-word Unicode component). This is the **resource-directory argument axis**, not a relocated compiler installation, and does not require headers to exist. - In both C and C++, the actual parent metadata differs: the spaced path leaks the literal bare word `Tools`. Candidate metadata is equal across those paths; no-space C/C++ bytes remain byte-identical to the parent's. Separate real CPU (`x86-64-v2`), feature (`-mno-avx2`), target (`x86_64-w64-windows-gnu`) and option (`-fno-stack-protector -fno-pic`) mutations remain distinct. The actual quoted nonpath macro value `PUBLIC_LABEL=hello wide world` survives. - The Linux run made **18 real driver calls plus one real `--version` call**, all native exit 0; harness exit 0. Two read-only Git blob reads and one `ldd` dependency discovery also exited 0 (22 subprocesses total). Complete argv, UTC timing, raw stdout/stderr, native return codes, stamp bytes and hashes, and the real parent differential are retained in the local `linux-driver` evidence packet. The recording adapter delegates to real subprocesses; it supplies no mock compiler response. - Production/test source, compiler, native Python executable and every resolved shared-library file were SHA-256-identical before/after. The compiler hash also matches its frozen host-659 manifest. `python -I -B` was used; child commands had isolated HOME/XDG/TMP, a process-scoped library search path and individual owned process groups with 20-second timeouts. The shared heavy lock was held. Before/after samples for every invocation stayed above 3 GiB disk and MemAvailable; these are samples, not continuous monitoring. All owned child groups were absent after exit, with an independent post-run process readback retained. - Windows Clang **23.1.5** and Linux Clang **18.1.3** are separately qualified within each OS. No cross-version or cross-OS equality of normalized bytes is expected or claimed. The earlier Windows 32 driver calls and two version calls are preserved unchanged and were not rerun. No source edit, full-suite rerun or frozen-baseline rerun was needed. `-###` does not compile, link, validate headers, invoke the complete preprocessing `stamp()`, run a game or use a GPU. **New physical Linux placement checks (2026-11-16):** On the same unchanged candidate and frozen parent, real ELF Clang 18.1.3 was independently copied into `NoSpaceClang` and `Trois Mots épreuve`, with 17 resolved libraries and 237 existing resource files per prefix (510 files, 487926162 bytes total). These are distinct physical executable/dependency/resource copies, not resource-argument substitutions; paired runs used identical cwd, environment and options, without a `-resource-dir` override. Across eight configurations, 32 real driver calls plus 11 other native calls all exited 0. C and C++ path baselines are candidate-equal, parent-RED and no-space byte-unchanged; C-only CPU, feature, target, default-option and quoted-value controls remain distinct. Source and original/copied tools were unchanged. The OS loader and kernel remain shared; copied resource/header existence is not header validation. **New physical Windows placement checks (2026-11-16):** Existing LLVM-mingw Clang 23.1.5's real wrapper, `clang-23.exe` backend, recursive DLL/config closure and resources were independently copied into no-space and three-word Unicode prefixes (924 files, 469882234 bytes total). The 11 actual config files and 10 same-folder includes were kept unchanged. Initial 32 driver calls plus four version/resource calls all exited 0; C and C++ path baselines are candidate-equal, parent-RED and no-space byte-unchanged, while C-only CPU, feature, target and quoted-value controls remain distinct. The original initial **PRODUCT_RED** is retained: two default-option oracle assertions expected `-fno-stack-protector -fno-pic` to change metadata, but the raw cc1 arguments were already identical before parsing. A separate additional C-only `-fstack-protector-strong` control made four real calls and confirmed meaningful raw and normalized differences at both placements, candidate path equality, parent RED and unchanged no-space bytes. It did not recopy tools, replay initial cases or modify original evidence/inputs. Combined Windows totals are 36 driver calls and 40 native wrapper invocations, not a claim about actual process count; the separate three benign Job controls are not product tests. **Scope and retained limits:** These new physical-placement checks are separate from the earlier resource-argument checks, not replacements or full-suite reruns. Linux's original global-process audit **FAIL** remains retained because some `/proc` entries were unreadable; independent SPEC and QUALITY adopted only the scoped evidence, including absence of all 43 owned child groups. Windows owned-Job and readable own-namespace readbacks were empty; there is no global host process-absence or runtime DLL-load-trace claim. The combined Windows assessment is scoped to these retained receipts, not maintainer approval or a full-suite pass. Neither OS result establishes installation/SDK completeness, full `stamp()`, transitive preprocessing, header usability, compilation, linking, object reuse or gameplay. Different Clang versions are qualified separately; no cross-OS/version metadata equality is claimed. The existing Linux full-suite exit 1, all 22 failures and strict-comparator false remain unchanged. **Retained harness history:** The first Linux placement harness stopped on a lexical `ldd`-path membership assertion; a scoped continuation reused the same verified physical copies and corrected only that bookkeeping check, without a product edit or recopy. Before Windows compiler execution, review corrected launch containment (suspended creation, owned-Job assignment, then resume) and lossless diagnostic-byte packaging; benign exit/nonzero/descendant-timeout controls exercised that harness separately. The first Windows inspection also stopped on same-folder config includes until the actual finite config closure was reviewed and pinned. The separate four-call control had a failure-path raw-packaging correction reviewed before execution, and an invocation without its required arming flag stopped before any native call. Those review, harness and fail-closed preflight receipts remain retained alongside the initial results; none is erased, counted as a compiler pass or used to change product code. **Not tested:** full `stamp()` preprocessing/header behavior, a native Windows build, shipped third-party-object reuse, or any game. Compiler replies in the original focused tests remain labelled synthetic fixtures; the separate native driver checks above are not mocked. **Must be tested:** - Windows and Linux, any GPU: by hand, validate complete `stamp()` preprocessing with usable headers and shipped third-party-object reuse, expecting unchanged inputs/options to select compatible metadata and objects and semantic changes to remain distinguishable. Physical driver placement is now exercised within the scoped checks above; that does not establish these remaining build-level behaviors. No test-farm request has been made. ## Risks **What could break:** parsing a Clang argument transcript unlike the quoted/escaped form covered here could still prevent reuse. The path filter's existing semantics are unchanged, not broadened to remove semantic options. **Turning it off:** revert the commit. **Compatibility:** no new flags, defaults, manifest format, runtime code, saves, renderer ABI or shader/pipeline format. Previously contaminated spaced-path metadata changes, intentionally, to the existing path-independent value; no-space metadata is tested byte-for-byte unchanged. ## Checklist ### Everyone - [x] One topic. - [x] The title is a Conventional Commit (`type(scope): description`) of the status line's type. - [x] Why gives the reason plainly and links the issues that fit (the public tracker and complete public PR file lists searched; parent scope adopted, no matching issue claimed). - [x] Measurements has a verdict, or one line saying why there is nothing to measure. - [x] Testing names what was tested, what wasn't, and what must still be tested, where and how. - [x] Rebased on current `main` at frozen intake (`d9486a0846e90905e2ee0bdc4a0d6df96c547c87`). Public main has since advanced; no rebase to that later tip is claimed. Relevant source closure is checked again before publication, without transferring full-suite results to another tree. - [x] `python -I -B tests/run_all.py` qualified against its exact pinned-main baseline: native exit 1 is preserved, all 22 existing failures are named above, and no new failure or changed skip status was found. This does not assert a wholly passing suite. - [x] No game files, dumps, build output or generated files. - [x] `third_party/` untouched (a new folder with a `PROVENANCE.md` instead). - [x] The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled. ### When it applies - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. - [x] A game-free test added for the fix, failing on `main` and passing here. - [x] A check under `tests/` that reads changed source updated in this PR, not weakened (new independent file). Source export scope: unchanged gpu/build-shadertool.bat has the recorded .gitattributes LF-to-CRLF conversion. Candidate production and test files match native Git blobs; source/modes are unchanged across each run. Tooling is absent in the isolated runner context: two corresponding failure messages require only an exact sibling-path substitution; the original strict comparison remains false. This is not hosted-CI or literal exact-tree equivalence. **Windows qualification caveats:** Native CPython 3.14.7 used a new independent copy of an existing pinned-dependency runtime, not a venv or a replacement product runner. Each candidate used its own full frozen-head public Git-blob export (2652 tree entries; 2651 files materialized; `AGENTS.md` kept opaque and never materialized). Production/test SHA-256 and all exported file bytes were verified before/after; Git modes remain metadata, not a claim of POSIX permission semantics on NTFS. Normal public test sources and commands were unchanged. Full stdout/stderr and exact native executable/argv/cwd/exit records are retained. A post-test harness assertion exited 1 because five CPython encoding bytecode cache files appeared only in the new owned runtime copy; independent readback confirmed no existing runtime file changed, the original runtime stayed byte-identical, and all test invocations exited 0. That earlier fixture run did not qualify a real compiler. The separate native driver-only checks above do; no full Windows suite, compilation, game, GUI, hosted CI or publication qualification is claimed. Existing Linux full-suite results above are unchanged and are not transferred to Windows.
fix(build): keep quoted compiler paths out of object stamps
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request_target) Has been skipped
PR description / Description follows the template (pull_request_target) Successful in 3s
2370dd31fd
Tokenize complete Clang -### arguments before filtering paths, without
interpreting Windows backslashes as shell escapes. Preserve CPU, target
and other compiler options. Add game-free fixture regression controls.

Assisted-by: Hermes Agent (OpenAI GPT-6.1-sol)
Co-Authored-By: Claudia (OpenAI GPT-6.1-sol)
NitnekLeBorgne changed title from WIP: fix(build): preserve quoted third-party driver paths to fix(build): preserve quoted third-party driver paths 2026-11-16 02:58:27 +00:00
Member

Code-only review of 2370dd31fd; informational findings, with final approval left to another reviewer.

No introduced correctness defect found in the two-file diff. Keeping each quoted argument together prevents words inside spaced installation paths from surviving the existing path filter. The subprocess invocation and cache-copy behavior are unchanged. The new tests exercise the actual _driver() parser using mocked compiler replies, including escaped quotes, Windows separators, Unicode, semantic values and failed/missing compiler output; restoring the old whitespace split would break the spaced-path regression.

The remaining validation is already identified in the description: capture real Clang -### output on Windows and Linux from equivalent installations with and without spaces, confirm equal metadata for identical options, and confirm a CPU/target change still changes it. Reported native Windows and Linux fixture passes do not establish that transcript coverage. No unrelated game run is needed for this parser change.

Read the full diff, surrounding stamp/cache code, existing tests and reported validation. No tests, builds, compiler invocations or game runs performed in this review. No unexplained script side effects found.

Code-only review of 2370dd31fdc1aa330dff666108cdadde3159984b; informational findings, with final approval left to another reviewer. No introduced correctness defect found in the two-file diff. Keeping each quoted argument together prevents words inside spaced installation paths from surviving the existing path filter. The subprocess invocation and cache-copy behavior are unchanged. The new tests exercise the actual `_driver()` parser using mocked compiler replies, including escaped quotes, Windows separators, Unicode, semantic values and failed/missing compiler output; restoring the old whitespace split would break the spaced-path regression. The remaining validation is already identified in the description: capture real Clang `-###` output on Windows and Linux from equivalent installations with and without spaces, confirm equal metadata for identical options, and confirm a CPU/target change still changes it. Reported native Windows and Linux fixture passes do not establish that transcript coverage. No unrelated game run is needed for this parser change. Read the full diff, surrounding stamp/cache code, existing tests and reported validation. No tests, builds, compiler invocations or game runs performed in this review. No unexplained script side effects found.
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Guide screenshots / Guide images and local UI source review (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request_target) Has been skipped
PR description / Description follows the template (pull_request_target) Successful in 3s
This pull request doesn't have enough approvals yet. 0 of 1 approvals granted.
Some workflows are waiting to be reviewed.
This branch is out-of-date with the base branch
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'fix/third-party-driver-paths-20261116:NitnekLeBorgne-fix/third-party-driver-paths-20261116'
git switch 'NitnekLeBorgne-fix/third-party-driver-paths-20261116'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!1021
No description provided.