fix(ci): kill the ItemsTest/chain-test exit-crash flake at its root — BlitzForge bump (#85 + #86)#549
Merged
Merged
Conversation
…ause fix Bumps the BlitzForge submodule to pick up blitz-forge#85 + #86: - #86 (the root cause of the long-standing ItemsTest / OnlinePlayerChainTest CI flake): deleteVars() codegen emitted a GC release for program GLOBALS through an uninitialized Decl::offset, making main's epilogue perform a wild [ebp+garbage] stack-relative read -- harmless when the stale value was 0, an intermittent access violation otherwise. Verified 0/300 + 0/300 post-fix vs a 2-12% baseline. - #85: Delete Each walk-safety (captured-next invalidation left chain remainders undeleted), GC reference_map erase on pool delete, and permanent native-crash diagnostics (exception code, module-relative fault address, registers, generated-image dump, symbolized stack walk). CLAUDE.md's 'Known intermittent flake' section is retired accordingly (the close/reopen retry ritual is no longer the protocol; diagnose from the new crash diagnostics if anything similar ever reappears). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
rockinraymond
pushed a commit
to rockinraymond/rcce2-mythic
that referenced
this pull request
Jun 11, 2026
Picks up blitz-forge#87: the quality-gate review of the flake fix (RydeTec#549 / blitz-forge#86) found the new seTranslator diagnostics sprintf'd a 110-char regs line into a 96-byte stack buffer -- a guaranteed 15-byte OOB write on the crash path. All handler formatting is now snprintf-bounded, and the generated-image dump clamps to the committed region it actually reads. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary (non-technical)
Every PR in this repo has carried a ~5-7% chance of a spurious CI failure for months:
ItemsTest.bborOnlinePlayerChainTest.bbwould crash after all tests passed, and the documented workaround was to close and reopen the PR to re-roll CI. The cause was never in those tests — it was a code-generation bug in the BlitzForge compiler. This PR bumps the submodule to the fixed compiler and retires the workaround from CLAUDE.md. CI failures mean something again.Root cause (in the compiler, fixed upstream)
blitz-forge#86:
deleteVars()emitted the GC release__bbRelease([ebp + Decl::offset], type)for every struct-typed decl in main's environ — including program globals, whoseDecl::offsetis never assigned by frame layout and was missing from theDeclconstructor's initializer list (uninitialized memory). Main's epilogue therefore executed a wild[ebp+garbage]read: harmless when the stale heap value was 0 (most runs), an access violation otherwise (the historical "Stack overflow!" label was a separate, since-fixed mislabeling bug). GC on/off irrelevant — the wild read precedes the release, which is why the non-GC chain test crashed identically.blitz-forge#85 (also in this bump): two real
Delete Eachruntime defects found en route (captured-next invalidation that silently left chain remainders undeleted — deterministic 499/1000-survivor repro, now a regression test; GC reference_map desync on pool delete), plus permanent native-crash diagnostics (exception code, module-relative fault address, registers, generated-image hex dump, dbghelp-symbolized stack walk) — the instrumentation that made the root cause findable.Evidence
All pre-fix failures:
0xC0000005at a stable program-relative offset in main's epilogue. At baseline rates ~6-12 and ~27-36 failures were expected across 300 runs; observing zero in both is p < 1e-8. Full investigation log:compiler/BlitzForge/docs/notes/exit-teardown-crash.md.This PR
af497fba(develop, includes ci: add GitHub Actions workflow (compile + test) #85 + tests: Environment TimeDelta + SafeWrite helpers (Track R, first wave) #86).compile.bat(engine + tools) clean andtest.bat50/50 green against the fixed compiler.Note: this PR's own CI run rebuilds BlitzForge (cache miss on the new submodule SHA) — expected to be slower than usual.
🤖 Generated with Claude Code