ci: the per-PR repsel matrix arms cannot reproduce the relocating-minor defect class (evac base is push-only)

[proggeramlug]

The hole

scripts/gc_repsel_matrix.sh runs on pull requests with --arms pr:

PR_ARMS="default,verify_evac,cons_scan_off,shipped_default"

and on push with --arms all (.github/workflows/test.yml, gc-stress job).

But the evacuating base that #6981 established as the discriminator needs both halves together:

PERRY_GC_INCREMENTAL=0 PERRY_CONSERVATIVE_STACK_SCAN=off

and that combination appears only in --arms all (evac_minor / cons_scan_off_force). cons_scan_off on its own is pressure + precise roots with incremental mode still on, which does not take the relocating path.

Measurement

Built the five #6982 crashers from aa1c15028 (i.e. unfixed) and ran each PR-gated arm. All green:

                        default  verify_evac  cons_scan_off  shipped_default
repsel_canonical_i32    0        0            0              0
typedarray_param_read   0        0            0              0

The same binaries crash immediately under PERRY_GC_HEAP_LIMIT=8 PERRY_GC_INCREMENTAL=0 PERRY_CONSERVATIVE_STACK_SCAN=off (SIGBUS/SIGSEGV, copied_objects 4 579 / 5 965).

So the entire "raw reference held across a relocating collection" class — #6951, #6972, #6982, #6991, #6992 — is invisible to the per-PR gate. It can only go red after merge, on push.

Why I did not just fix it

Adding evac_minor to PR_ARMS today would fail every PR: 13 of 21 corpus cells are still red on that arm (2 crashes + 11 mismatches after #6982's fix). The gate has to become green-able first.

Suggested shape

1. Land the outstanding rooting fixes (fix(gc): operand temporaries are precise roots — collection methods, constructor args, string-method receiver (#6970, #6969, #6971) #6983, gc: compiled constructor receiver goes stale across the globalThis-population collection (unmasked by #6982) #6991, gc: proven-this method value is lost across a relocating minor (TypeError: bump is not a function) #6992, gc: scalar-replaced object/array locals holding heap values are not precise roots #6968) until the evac_minor column is green or has a documented, enumerated skip list.
2. Then add evac_minor to PR_ARMS, with the remaining known-red cells in an explicit allowlist (same provenance discipline as test-parity/known_failures.json, cf. test-parity: audit known_failures.json — every entry needs an issue # and date #797) so new breakage is caught but known breakage does not block.
3. Until then, consider a non-blocking evac_minor PR run so the signal is at least visible on the PR page rather than only after merge.

Related: #6981, #6982, #6991, #6992, #6977, #6950.

[proggeramlug]

Datapoint for this issue's thesis, from a GC measurement run (#7056): the relocating-minor defect class reproduces in the shipped default configuration, with no special arm at all.

#7055 is an ordinary event-loop TypeScript program that produces a deterministic wrong checksum under plain perry file.ts -o bin && ./bin. The discriminator is exactly copied_objects > 0:

arm	minors @600	copied	checksum
shipped default	56	11,661,882	wrong
PERRY_GC_SCAVENGE_NURSERY_MB=128	3	0	correct
PERRY_GC_MOVING_LOOP_POLLS=0	1	0	correct
PERRY_GC_MOVING_LOOP_POLLS=0 PERRY_GC_SCAVENGE=1	60	11,351,683	wrong

Note the second row: three minor cycles, right answer, because none of them relocated.

This bears on the "suggested shape" here in two ways.

The gate hole may be wider than evac_minor. This issue frames the missing coverage as the PERRY_GC_INCREMENTAL=0 PERRY_CONSERVATIVE_STACK_SCAN=off combination that only --arms all reaches. But the case above needs neither half — it is the default. PERRY_GC_TRACE=1 on the shipped arm shows root_sources.native_stack_fallback.decision = "skip_disabled" on 62 of 63 cycles, so the shipped moving path is already running with precise roots and no conservative scan. #6981's reasoning that gc_check_trigger's forced force_full_scan() keeps this class out of production does not cover the deferred/moving path.

What a green-able discriminator might key on. Since the discriminator here is copied_objects > 0 rather than any particular arm, a cheaper gate than adding evac_minor to PR_ARMS may be available: run the corpus in the default configuration but fail if a test that relocated disagrees with a copied_objects == 0 run of the same binary. That is a self-checking property (no oracle needed beyond Perry-vs-Perry), it does not require the 13-of-21 red cells to be fixed first, and it would have caught #7055 on the PR that introduced it.

Full measurement context, including the per-cycle trace figures quoted above: #7056.