gc: two rooting gap tests crash on macOS under PERRY_NO_AUTO_OPTIMIZE=1 while Linux passes them (untriaged)

[proggeramlug]

Found in passing during #7599's known-failures audit, on the shared dev Mac
(macOS arm64, perry-dev profile, current main of 2026-08-08):

test_gap_gc_rest_argument_rooting and
test_gap_gc_same_module_call_argument_rooting crash on macOS when compiled
with PERRY_NO_AUTO_OPTIMIZE=1, while Linux (the gap_snapshot baseline
platform) records both as passing.

Not yet triaged — this issue exists so the observation is not lost. Three
things whoever picks it up should know:

1. These are gc rooting acceptance tests (the gc(layer 3): from-space quarantine catches 55 stale dereferences across the gap suite — the instrument is in CI but aimed at one synthetic fixture #7341 family), so a platform-
   split crash is potentially a real macOS-only rooting hole, not a harness
   quirk — the auto-opt and no-auto-opt runtimes differ in feature set and
   link layout, and the no-auto-opt arm is the one ordinary perry compile
   users get when the auto-opt path is bypassed.
2. test_gap_gc_same_module_call_argument_rooting has crashed spuriously
   before — during a disk-full window on 2026-08-07 it recorded a crash that
   re-ran 10/10 clean afterwards (fix(runtime): a Map/Set subclass in a base-typed binding was read as a raw header (#7570) #7573's report). First step is therefore
   reproduce N/N on a healthy machine before believing it; the test(parity): make known_failures.json a ratchet, not a suppression list (#7582) #7599 agent
   saw it in passing, not under controlled conditions.
3. If it reproduces, the instruments are PERRY_GC_ZEAL=1 PERRY_GC_PROTECT_FROMSPACE=1 PERRY_GC_DIAG=1 — but note gc: zeal + compiled loop polls performs zero copying minors on compute-only benches — green instrument runs may be vacuous #7604: on
   compute-only shapes the zeal recipe can arm without ever firing, so check
   retired_set= lines actually appear before trusting a clean run.