gc-stress: test_gap_repsel_gc_stress fails on every full mark-sweep arm since ~2026-09-03 (suspect: aa8d2ade0 stops marking the remembered set on full traces) #9782

gc-stress matrix (4/4) has been red on main since roughly 2026-09-03. It is a
deterministic failure — byte-identical counters across every run and every SHA I
checked — and the arm split points at a full-collection root-scanning bug, i.e. a
potential collected-live-object class of defect.

The failure

run [8/20] test=test_gap_repsel_gc_stress env=_base arm=gen_gc_off
FAIL test_gap_repsel_gc_stress (output-mismatch cycles=43 evacuated=0 scavenged=0 reclaimed=1327926376)
run [8/20] test=test_gap_repsel_gc_stress env=PERRY_WRITE_BARRIERS_0 arm=wb_off
FAIL test_gap_repsel_gc_stress (output-mismatch cycles=43 evacuated=0 scavenged=0 reclaimed=1327926376)
run [8/20] ... arm=gen_off_verify   -> FAIL (same signature)
run [8/20] ... arm=wb_off_force     -> FAIL (same signature)
run [8/20] ... arm=all_four         -> FAIL (same signature)

repsel_gc_stress  PASS PASS PASS PASS PASS PASS FAIL FAIL FAIL FAIL FAIL PASS PASS PASS PASS PASS PASS PASS PASS PASS PASS PASS
summary: PASS=360 UNVER=75 XFAIL=0 FAIL=5

Example run: https://github.com/PerryTS/perry/actions/runs/33829521137 (job 100904513692).
Identical on later runs, e.g. https://github.com/PerryTS/perry/actions/runs/33926006467.

The discriminating evidence: which arms fail

Exactly five arms fail — gen_gc_off, wb_off, gen_off_verify, wb_off_force,
all_four — and those are precisely the arms that disable generational GC or write
barriers, forcing a full / mark-sweep collection instead of a generational minor. All
fifteen other arms pass, including default, evac, force-evacuate and every
representation-selection off-switch.

evacuated=0 scavenged=0 with a large reclaimed=1327926376 is consistent with a
non-moving full sweep reclaiming substantially more than it used to.

Window

Every push between those two had its CI run cancelled (superseded) before any job started,
so there is no direct evidence for the intermediate SHAs. That leaves a 17-commit range,
dae0ae34c8cf..d239929b4c46.

Prime suspect — correlation only, NOT a proven bisection

aa8d2ade0 — "fix(gc): stop two passes retaining unreachable objects (#9628, #9629)"
(2026-09-03 22:01) is the only commit in the range that touches GC-cycle root-scanning
semantics. In crates/perry-runtime/src/gc/cycle.rs it makes the remembered set a root
source only for minor collections:

let mark_from_set = self.minor_only; // #9629: only a MINOR needs old->young
                                     // edges as roots; a full trace reaches
                                     // them through the live old objects
                                     // themselves...

So on any non-minor (full) collection the remembered set is no longer marked as roots —
which is exactly the code path the five failing arms force. test_gap_repsel_gc_stress is
purpose-built to hold representation-selected locals (Ptr<Shape>-proven objects and
friends) live across escaping churn, which is the shape that would break if such an object's
old-gen reachability depended on the remembered-set path now skipped on full traces.

The very next commit, 51bfe741d ("test(gc): the remembered-set slicing tests drive a
minor, not a full cycle"), rewrote two unit tests to match the new behaviour — so the
reduced retention was intended, but this stress fixture suggests an edge case those unit
tests do not cover.

Stated plainly: this is a code-review correlation, not a bisection. No CI run exists at
aa8d2ade0 itself, so it is not isolated from its 16 neighbours in the range. I did not run
the fixture at aa8d2ade0 and its parent because scripts/gc_repsel_matrix.sh requires a
full compiler build (cargo build --profile <p> of perry plus the runtime/stdlib
staticlibs) before it can compile and run the corpus — two of those plus two matrix runs is
well beyond a cheap check.

Correction: an earlier triage of mine attributed this to the #9644 idle-compaction
commits (3812cc973, 42cd82fdf). That was wrong — both land after d239929b4c46,
which was already red. Please do not chase them.

Paths

• Fixture: test-files/test_gap_repsel_gc_stress.ts
• Harness: scripts/gc_repsel_matrix.sh (computes the cycles/evacuated/scavenged/reclaimed line and the arm matrix)
• Tier driver: scripts/release_sweep_tiers/tier04_gc_stress.sh
• Suspect code: crates/perry-runtime/src/gc/cycle.rs
• Suspect's own tests: crates/perry-runtime/src/gc/tests/retention_9628_9629.rs

Suggested next step

Build at aa8d2ade0 and at its parent 658007137, run
scripts/gc_repsel_matrix.sh restricted to test_gap_repsel_gc_stress and the
gen_gc_off arm, and compare. That converts the correlation into a bisection in two
builds. If it reproduces, the question is whether a full trace can in fact always reach
old->young targets without the remembered set when the young half was already swept.

https://claude.ai/code/session_01YPfnmWZmSpSWpmnoXvH8z2