§3.A follow-up — diag-arena/interner heap-state leak: bisection report

Investigation date: 2026-05-14 Worktree: agent-aa09a455338a982d3 (base commit 77c89a4d fix(ts-parity): tighten exact directive diagnostics) Status: Could not reproduce. The leak documented in the source NOTE block at packages/ts_conformance/src/ts_conformance.zig:2844 and in the journal entry "2026-05-14 — Phase 6 §6.A.4 follow-up: shared HOME_TS_CONFORMANCE_ROOT path discovery" no longer manifests on this branch.

Reproduction harness

Re-added an opt-in always-on test conformance: bisect exact-baseline heap leak placed BEFORE the adjacent unit tests it was claimed to corrupt (Zig executes tests in declaration order, so only a leading bisect can affect later tests). Toggle with HOME_TS_CONFORMANCE_BISECT=1; tune the slice with HOME_TS_CONFORMANCE_BISECT_{START,LIMIT} (defaults: START=0, LIMIT=25).

Source: packages/ts_conformance/src/ts_conformance.zig:2421

Bisection log

All runs use HOME_TS_CONFORMANCE_BISECT=1 plus the listed LIMIT. Adjacent tests checked: conformance: type-error decl fails as expected (line 2497) and conformance: runCorpus supports exact diagnostic entries (line 2621). All test counts are out of 45/45.

HOME_TS_CONFORMANCE_FULL=1 HOME_TS_CONFORMANCE_EXACT=1 HOME_TS_CONFORMANCE_LIMIT=50 also runs cleanly to completion (45/45). No expected 1, found 0 or expected .passed, found .failed failures observed in any configuration.

Hypothesis: the leak was already fixed

The journal entry that documented the leak landed at HEAD e46d293 (or thereabouts) on 2026-05-14. The current worktree base is six commits ahead at 77c89a4d. Two of those intervening commits touch packages/ts_checker/src/check.zig in ways that look directly relevant:

• f84d9ad0 fix(ts-parity): close conformance tail crashes — adds a new reentrancy guard resolving_exported_type_decls: std.AutoHashMapUnmanaged(NodeId, void) (check.zig +468 lines). Circular module references could legally mention an exported class/interface before its body was available; nested lookups now fall back to any instead of recursively rechecking the same declaration. This is a strong candidate for the cross-test corruption: a recursion that escaped the Checker instance lifetime would write into the diag arena / interner of a re-entrant invocation.
• dca93374 perf(ts-parity): clear full corpus coarse gate — additional check.zig changes (+42 lines); commit message claims it cleared the coarse gate to 100%, suggesting it removed a class of late-tail crashes.

Neither commit directly mentions the heap-state interaction, but the timing and the surface area lines up.

Suggested next steps for the fix agent

1. Don't chase a fixture that no longer breaks. Confirm independently on a clean build (./pantry/.bin/zig build test -Dfilter=ts_conformance --summary all — should report 45/45 in ~30 s) that the leak is truly gone.
2. Re-run the bisect harness (HOME_TS_CONFORMANCE_BISECT=1 HOME_TS_CONFORMANCE_BISECT_LIMIT=2000 ./pantry/.bin/zig build test -Dfilter=ts_conformance) before declaring the §3.A follow-up closed. If it stays green, retire the NOTE block at ts_conformance.zig:2844 and the §3.A entry in TS_PARITY_PLAN.md.
3. If it surfaces again, the bisect harness is already wired to print [bisect] RUN <idx> <name> for each case — read backwards from the first crashing adjacent test to find the offender. The most likely re-introduction vectors are: (a) new code in check.zig that mutates Checker-owned state past the Compilation.deinit() boundary, (b) interner-shard work in interner.zig (currently dirty in the main repo with broken references to shardIndexFor / Shard) being merged without finishing the API, or (c) any helper that returns slices borrowed from a per-Checker arena into a static / cross-test cache.
4. Promote the bisect to always-on once a couple more PRs land cleanly. The opt-in gate (if (!envBoolOne("HOME_TS_CONFORMANCE_BISECT")) return;) is the single line to remove. Recommended default LIMIT for an always-on gate: 25 — that's the original threshold that was claimed to corrupt and is short enough to keep CI under 30 s.