Retry the exact failed compile job #1053

Merged

monsterlag merged 1 commit from fix/retry-exact-compile-job into main

2026-11-09 15:15:28 +00:00

monsterlag commented

2026-11-09 08:12:31 +00:00

(Migrated from github.com)

Keep each completed future paired with its original job, retain that complete job on failure, and retry it once serially. compile_one's return API, memory/worker/cache policies, progress records and linking stay unchanged.

Refs #1029 as related investigation context; this is separate from the memory-reservation lease fix.

Validation:

Five authored cases execute the actual AST-extracted compile_one and compile/retry coordinator using a real ThreadPoolExecutor and fake compiler children. They cover different outputs/flags for one source, B's successful retry, persistent B failure returning1 before linking, both jobs failing, successful jobs without retries, and existing counted progress.
Unmodified main29571064 fails four cases and passes the success control. The preserved baseline reports recovery with B.o absent; the fixed source passes all five on Windows and WSL Python.
Windows strict automatic runner passes the new check and the existing compiler-failure check; git diff --check passes.

No native compiler, game, memory stress or full build/link. The fixture disables the memory gate and spies on cache publication/ETA helpers to isolate job selection. Existing stamp-file ResourceWarning diagnostics remain visible and are not suppressed.

All hosted checks on 502544b0 pass: Linux source regressions, Windows source regressions, Windows Vulkan renderer, Builder packaging, AppImage packaging, SPIR-V checks and workflow lint. Release publication was intentionally skipped.

Release integration: the official v2.0.8-rc.2 and rc.3 history contains this fix through merge commit 2c5c9e6. GitHub's PR metadata remains open while the default main branch is at a different revision; this note records released-source incorporation separately from a formal PR merge.

When two compile jobs share a source but have different flags and outputs, the retry pass currently finds the first job with that source. If A succeeds and B fails, it retries cached A and can announce recovery/proceed to linking while B's object is still missing; the linker may then fail. Two failures can likewise retry A twice. Keep each completed future paired with its original job, retain that complete job on failure, and retry it once serially. `compile_one`'s return API, memory/worker/cache policies, progress records and linking stay unchanged. Refs #1029 as related investigation context; this is separate from the memory-reservation lease fix. Validation: - Five authored cases execute the actual AST-extracted `compile_one` and compile/retry coordinator using a real `ThreadPoolExecutor` and fake compiler children. They cover different outputs/flags for one source, B's successful retry, persistent B failure returning1 before linking, both jobs failing, successful jobs without retries, and existing counted progress. - Unmodified main29571064 fails four cases and passes the success control. The preserved baseline reports recovery with B.o absent; the fixed source passes all five on Windows and WSL Python. - Windows strict automatic runner passes the new check and the existing compiler-failure check; `git diff --check` passes. No native compiler, game, memory stress or full build/link. The fixture disables the memory gate and spies on cache publication/ETA helpers to isolate job selection. Existing stamp-file `ResourceWarning` diagnostics remain visible and are not suppressed. All hosted checks on `502544b0` pass: [Linux source regressions](https://github.com/NotRexed/Hookshot/actions/runs/42584275577/job/129050210560), [Windows source regressions](https://github.com/NotRexed/Hookshot/actions/runs/42584275577/job/129050210473), [Windows Vulkan renderer](https://github.com/NotRexed/Hookshot/actions/runs/42584275563/job/129050209669), [Builder packaging](https://github.com/NotRexed/Hookshot/actions/runs/42584275563/job/129050210079), [AppImage packaging](https://github.com/NotRexed/Hookshot/actions/runs/42584275601/job/129050210246), SPIR-V checks and workflow lint. Release publication was intentionally skipped. Release integration: the official v2.0.8-rc.2 and rc.3 history contains this fix through merge commit [2c5c9e6](https://git.magicbean.dev/MagicBeanDev/Hookshot/commit/2c5c9e6). GitHub's PR metadata remains open while the default main branch is at a different revision; this note records released-source incorporation separately from a formal PR merge.

Commenting is not possible because the repository is archived.

No reviewers