bun test: exit 1 cleanly on failing runs under LeakSanitizer

[robobun]

Fixes #32183. Stacked on #32180 (base branch is farm/1b909a84/test-exit-lsan-leak); only the last commit is this PR. It needs that PR's collect_for_leak_check_at_exit(), so it should land after #32180 and be rebased onto main at that point.

Problem

A failing bun test run aborts (SIGABRT, exit 134) instead of exiting 1 when ASAN_OPTIONS=detect_leaks=1:abort_on_error=1 is set and BUN_DESTRUCT_VM_ON_EXIT is not, which is exactly how the ASAN CI lane runs subprocesses. #32180 fixed this for green exits; failing exits abort for three more reasons:

1. --bail on an in-test failure calls Global::exit(1) from CommandLineReporter::handle_test_completed while the active file is mid-execution, skipping all teardown. No collection runs, and the run loop's strong ref plus live BunTest::run frames mean a collection at that depth could not free the test-runner boxes anyway.
2. The module-evaluation-failure bail branch in TestCommand::run tears down inline (deinit_for_exit(), RUNNER clear) but exits without the final collection, leaving Expect/RefData/BunTestCell boxes malloc-live.
3. Any failing run (bail or not): printing the failure remaps the stack trace, which caches a ParsedSourceMap Arc in vm.saved_source_map_table via SavedSourceMap::get_with_content. The table stores entries as tagged pointers (Value), which LSan does not recognize as pointers, so the live cache entry is reported as a leak:

Direct leak of 168 byte(s) in 1 object(s) allocated from:
    <alloc::sync::Arc<bun_sourcemap::parsed_source_map::ParsedSourceMap>>::new
    <bun_jsc::saved_source_map::SavedSourceMap>::get_with_content  src/jsc/SavedSourceMap.rs:365
    <bun_jsc::virtual_machine::VirtualMachine>::remap_zig_exception

Fix

• New TestRunner.bailed flag. Both bail sites print the summary, bail message, and junit report exactly as before, then set the flag instead of exiting. step_sequence_one starts nothing further once the flag is set (no test callbacks, no remaining hooks, matching the old hard-exit behavior), BunTest::run discards queued results so nothing reports after the bail message, the drive loop in TestCommand::run stops ticking the event loop and returns, and run_all_tests skips remaining files. TestCommand::exec then exits 1 through the same teardown sequence as the normal tail (deinit_for_exit, RUNNER clear, collect_for_leak_check_at_exit, global_exit), skipping the tail's snapshot/coverage/summary work the old hard exit also skipped.
• New SavedSourceMap::clear() releases every cached entry in place (same release_value the Drop impl uses, table stays valid and empty), called from collect_for_leak_check_at_exit(). This covers serial runs, both bail paths, and --parallel workers, whose exit path already calls that method. No-op in non-ASAN builds, which keep the fast exit.

The load-bearing lines are the this.jest.bailed = true assignments, the step_sequence_one early return, and the source_mappings.clear() call; the rest is the unwind plumbing.

Bail behavior is unchanged: the message prints at the same point, nothing runs after it (including afterAll, whose absence under bail is pre-existing and tracked in #12250), remaining files, repeats, and reruns are skipped, junit is still written, and exit code is 1. --bail is still not forwarded to parallel workers, and the coordinator's own bail path is untouched.

Verification

New failing exits are leak-clean under LeakSanitizer tests in test/cli/test/isolation.test.ts (serial failing run, --parallel failing run, --bail mid-file, --bail on module evaluation failure), gated on isASAN && isLinux like the green-run cases from #32180. All four fail on the unfixed build (exit 134 with the SavedSourceMap::get_with_content leak above, plus Expect::call/BunTest::ref_ frames for the bail cases) and pass with the fix. The bail cases also assert the bail contract: Bailed out after 1 failure, nothing runs after the bail, Ran 1 test across 1 file..

Also verified on a debug ASAN build:

• existing bail coverage passes: bun-test.test.ts (79 tests), regression/issue/12250.test.ts, regression/issue/26851.test.ts (junit outfile on bail), parallel.test.ts bail test
• full test/cli/test/isolation.test.ts passes (26 tests)
• no hang and no straggler output when --bail fires with concurrent tests still in flight (exits 1 in ~0.6s with 30s timers pending)
• the bail exit is clean both with BUN_DESTRUCT_VM_ON_EXIT=1 (a path the old hard exit never reached) and without
• --bail --isolate, --bail --rerun-each=3 behave as before

The test/no-validate-leaksan.txt exemptions for bail-spawning tests are left in place; dropping them is follow-up once the lane proves stable, since those files also exercise unrelated subprocess failure modes.

[robobun]

^{Updated 9:58 PM PT - Jun 20th, 2026}

❌ @robobun, your commit 6f1f108 has 2 failures in Build #63718 (All Failures):

• test/cli/install/migration/complex-workspace.test.ts - code 1 on 🐧 25.04 aarch64
• test/js/node/zlib/zlib.test.js - crash reported on 🐧 13 x64-asan

---

🧪   To try this PR locally:

bunx bun-pr 32187

That installs a local version of the PR into your bun-32187 executable, so you can run:

bun-32187 --bun

[claude]

I didn't find any bugs, but this restructures --bail from a hard Global::exit(1) into a flag-and-unwind across the test runner's run loops, plus a new SavedSourceMap::clear() on the exit path — worth a human pass to confirm the unwind fully preserves bail semantics in all the paths it now threads through.

Extended reasoning...

Overview

This PR changes how bun test --bail exits so that failing runs are leak-clean under LeakSanitizer. It touches 7 Rust source files plus tests:

• src/jsc/SavedSourceMap.rs: new clear() that releases all cached tagged-pointer entries in place.
• src/jsc/VirtualMachine.rs: collect_for_leak_check_at_exit() now calls source_mappings.clear().
• src/runtime/test_runner/jest.rs: new TestRunner.bailed flag.
• src/runtime/cli/test_command.rs: both bail sites (handle_test_completed and the module-eval-failure branch in run) now set bailed = true instead of calling Global::exit(1) / inline teardown; exec gains a centralized bail-exit block; the drive loop and run_all_tests check the flag to stop and unwind.
• src/runtime/test_runner/Execution.rs: step_sequence_one early-returns Done once bailed is set.
• src/runtime/test_runner/bun_test.rs: BunTest::run discards queued results once bailed is set.
• src/runtime/cli/test/parallel/runner.rs: comment-only update.
• test/cli/test/isolation.test.ts: hoists leakCheckEnv and adds 4 new LSan-gated tests for failing/bail exits.

Security risks

None. This is internal test-runner exit-path plumbing and a source-map cache teardown helper; no auth, network, parsing of untrusted input, or permission surfaces are touched.

Level of scrutiny

Moderate-to-high. The core change replaces an immediate exit(1) mid-execution with a cooperative unwind that must be observed at every layer (step_sequence_one, BunTest::run's result drain, the tick()/drive loop in TestCommand::run, run_all_tests, and finally TestCommand::exec). Any layer that misses the flag could run extra tests/hooks, print after the bail message, or hang on a still-ticking event loop. The new bail-exit block in exec also re-implements the teardown sequence with raw-pointer reborrows of vm inside run_with_api_lock. The PR description is thorough and lists the existing bail test files that were re-run, but this is exactly the kind of control-flow refactor where a second pair of eyes on the unwind ordering and the SavedSourceMap::clear() lock/iteration is worthwhile.

Other factors

• Stacked on #32180 (base branch farm/1b909a84/test-exit-lsan-leak); needs to land after that and be rebased.
• The new tests are gated on isASAN && isLinux, so most CI lanes won't exercise them; the bail-semantics assertions (Bailed out after 1 failure, no MARKER_AFTER_BAIL, Ran 1 test across 1 file.) are good but only run on the ASAN lane.
• Bug-hunter found nothing; CI build was just kicked off when I reviewed.

[claude]

I didn't find any bugs, but this replaces --bail's hard Global::exit(1) with a flag-based unwind across several layers of the test runner's run loop — worth a human pass to confirm the unwind plumbing covers all the edge cases (concurrent groups, in-flight hooks, the stacked-PR ordering with #32180).

Extended reasoning...

Overview

This PR makes failing bun test runs exit 1 cleanly under LeakSanitizer instead of SIGABRT'ing. It touches 7 source files and 1 test file:

• src/runtime/test_runner/jest.rs: adds a bailed: bool flag to TestRunner.
• src/runtime/cli/test_command.rs: replaces both --bail hard-exit sites with bailed = true, adds a centralized bail-exit teardown block in exec(), and threads bailed checks through the drive loop and run_all_tests.
• src/runtime/test_runner/Execution.rs: step_sequence_one early-returns Done once bailed is set so no further callbacks/hooks start.
• src/runtime/test_runner/bun_test.rs: BunTest::run discards queued results once bailed is set so nothing reports after the bail message.
• src/jsc/SavedSourceMap.rs: new clear() that releases all cached tagged-pointer entries in place.
• src/jsc/VirtualMachine.rs: collect_for_leak_check_at_exit() now also calls source_mappings.clear().
• src/runtime/cli/test/parallel/runner.rs: comment-only update.
• test/cli/test/isolation.test.ts: hoists leakCheckEnv and adds 4 new ASAN-gated tests for failing exits.

Security risks

None. This is internal test-runner control flow and exit-path teardown; no auth, network, parsing of untrusted input, or permission surfaces are touched.

Level of scrutiny

This warrants careful human review. The core change — converting --bail from a synchronous Global::exit(1) (which guaranteed nothing else ran) into a cooperative flag-based unwind across step_sequence_one → BunTest::run → TestCommand::run drive loop → run_all_tests → TestCommand::exec — is a non-trivial control-flow refactor of the test runner's hot path. The hard exit was a blunt but airtight guarantee; the unwind must now correctly stop every execution path (concurrent groups mid-flight, hooks, queued result-queue items, event-loop ticks, remaining files/reruns) without hanging or printing stragglers. The PR description shows thorough manual verification (concurrent tests with 30s timers pending, --bail --isolate, --bail --rerun-each, BUN_DESTRUCT_VM_ON_EXIT=1), and the new tests assert the bail contract, but a human should sanity-check the unwind coverage and the new SavedSourceMap::clear() lock/release sequence.

Other factors

• Stacked PR: this is stacked on #32180 and depends on its collect_for_leak_check_at_exit(); merge ordering matters.
• Test coverage: the 4 new tests are gated on isASAN && isLinux, so they only run in one CI lane. The bail-contract assertions (Bailed out after 1 failure, no MARKER_AFTER_BAIL, Ran 1 test across 1 file.) are good, and the author reports existing bail regression tests (bun-test.test.ts, #12250, #26851, parallel.test.ts) pass.
• SavedSourceMap::clear(): mirrors the existing Drop impl's release loop but leaves the table valid; correctly invalidates find_cache/last_ism first and is called only on ASAN exit paths after the final GC, so the risk surface is narrow.
• No prior reviews on this PR; CI build was triggered but status not yet reflected in the timeline.

[robobun]

CI status: every lane is green except one debian 13 x64-baseline test job, which failed on test/js/web/streams/streams-leak.test.ts ("native ReadableStream reuses the pull buffer across small reads", expected >20% reuse, got 9-18% across 4 in-job retries). That test is known flaky under CI load (wire coalescing changes the measured ratio) and has two open de-flake PRs: #32192 and #31355. Nothing in this diff touches streams, and the ASAN-only exit code here is compiled out in that release lane. The darwin jobs are still queued behind the runner backlog.

Ready for review; a retry of that one job should clear the red.

[robobun]

Rebased onto the current farm/1b909a84/test-exit-lsan-leak tip (#32180 was re-based onto a newer main, so its commit SHAs changed out from under this branch). Replayed the single commit with git rebase --onto; no content conflicts, the diff is byte-for-byte identical (212 insertions / 39 deletions across the same 8 files). The reported conflict was purely the divergent base history.

Re-verified on a debug ASAN build after the rebase: the 4 failing-exit leak tests and the 3 green-exit leak tests pass (7/7), and the --bail regression tests in bun-test.test.ts pass (5/5). Still stacked on #32180; merge order unchanged.

[claude]

No bugs found, but this refactors --bail from a hard Global::exit(1) into a flag-based unwind across several run-loop layers — worth a human eye on the unwind path (especially with concurrent tests in flight) and the stacked merge order with #32180.

Extended reasoning...

Overview

This PR makes failing bun test runs exit cleanly (exit 1, not SIGABRT) under LeakSanitizer. It touches 8 files: a new SavedSourceMap::clear() called from collect_for_leak_check_at_exit(), a new TestRunner.bailed flag, and the replacement of two Global::exit(1) bail sites with a flag-based unwind threaded through step_sequence_one, BunTest::run, the TestCommand::run drive loop, run_all_tests, and a new consolidated bail-exit block in TestCommand::exec. New ASAN+Linux-gated tests cover serial failing, --parallel failing, --bail mid-file, and --bail on module-eval failure.

Security risks

None. No auth, network, or untrusted-input handling. The change is internal test-runner control flow and exit-time memory cleanup that only affects ASAN builds (collect_for_leak_check_at_exit is a no-op otherwise).

Level of scrutiny

Moderate-to-high. The SavedSourceMap::clear() addition is straightforward (mirrors the existing Drop impl), and the test additions are well-scoped. But replacing a hard exit(1) with a cooperative unwind across five layers (handle_test_completed → step_sequence_one → BunTest::run → drive loop → run_all_tests → exec) is a meaningful control-flow refactor of user-visible --bail semantics. The old behavior was trivially correct (nothing runs after exit); the new behavior depends on every layer correctly observing the flag and short-circuiting without hanging, double-reporting, or running stragglers. The PR description reports manual verification of these cases (concurrent tests in flight, --isolate, --rerun-each, BUN_DESTRUCT_VM_ON_EXIT=1), and existing bail regression tests pass, but this is exactly the kind of change where a second pair of eyes on the unwind invariants is valuable.

Other factors

• Stacked PR: depends on #32180 for collect_for_leak_check_at_exit(); merge order matters and the base branch is not main.
• The step_sequence_one early-return (Done on bail) means in-flight concurrent sequences are abandoned without decrementing remaining_incomplete_entries; the unwind relies on the outer drive loop's !bailed guard rather than the phase reaching Done. This looks correct given the new return Ok(()) after the drive loop, but it's a subtle dependency between layers.
• No prior reviews from claude[bot] on this PR; CI is green except a known-flaky streams test unrelated to this change.

[robobun]

Thanks for the pass. On the concurrent-sequence point: that's correct and intentional. Once bailed is set, step_sequence_one returns Done so no further entries start, but the unwind does not rely on the phase reaching Done or on remaining_incomplete_entries hitting zero. The drive loop in TestCommand::run breaks on !bailed and return Ok(())s immediately, and BunTest::run discards any already-queued results, so in-flight concurrent sequences are simply abandoned (exactly as the old Global::exit(1) abandoned them). I verified this with concurrent tests still sleeping on 30s timers when the bail fires: the process exits 1 in ~0.6s with no hang and no straggler output.

CI status for this build: the new failing-exit leak tests pass on the x64-asan lane (19 of 20 asan shards green). The two red lanes are unrelated flakes:

• debian 13 x64-asan crashed in test/js/node/zlib/zlib.test.js with panic: N JavaScript functions were called outside of the microtask queue. That is a #[cfg(debug_assertions)] assertion in src/runtime/dispatch.rs and reproduces on a clean debug build of current main (b0fb1f767e) with none of this PR's changes (bun bd test test/js/node/zlib/zlib.test.js). It is newly visible on the release-asan lane because main recently enabled Rust debug-assertions there; this PR does not touch dispatch.rs, the event loop, or zlib, and that file never bails.
• ubuntu 25.04 aarch64 failed test/cli/install/migration/complex-workspace.test.ts with install failed / iconv-lite does not exist, a bun install registry fetch flake unrelated to the test runner.

Still stacked on #32180; merge that first, then this.