gc: shipped Perry runs at ~3x the RSS it needs — the nursery cap has never been on by default

[proggeramlug]

Summary

Shipped Perry carries ~3× the RSS it needs to. The 16 MB nursery cap that #7056 identified as the entire footprint win is gated behind two knobs that both default to off, so it has never been active in a shipped build.

Measured on today's collector (statepoints default, 148f97bc2), 8 gc_ratchet probes, as a 2×2 — because the one-armed version of this measurement says something false:

	no scavenge	scavenge
no cap (ships today)	799,604,736 B	799,604,736 B (+0%)
cap 16 MB	537,165,824 B (−33%)	245,006,336 B (−69%)

Wall time over the same probes:

arm	RSS	wall
ships today	baseline	2017 ms
cap only	−33%	2488 ms (+23%)
cap + scavenge	−69%	2085 ms (+3%)
scavenge only	+0%	2066 ms (+2%)

Read the row and the column, not one cell. Scavenge alone buys nothing — the top row is identical to the byte. The cap alone trades a third of the footprint for a quarter of the wall time. Only together are they worth it: the cap makes collections frequent, and scavenge makes them evacuating (O(live) copying) rather than O(heap) sweeps, so frequency becomes cheap.

This is why #7056's recommendation — "decouple the 16 MB cap and keep it" — is not on its own sufficient. Acting on it literally ships the −33%/+23% arm.

Why now

Enabling scavenge also defers alloc-point collections to a precise safepoint instead of collecting behind a forced conservative scan. That was unreasonable while precise roots were opt-in; it became reasonable when #7370 made native roots the default.

What blocks it

I prototyped the change (cap unconditional + gc_scavenge_enabled() default-on) and it is not landable as-is:

• 11/11 gc_ratchet probes stay byte-identical to Node, and the gap suite tracked the 447/19/13 baseline cleanly for the 200+ tests measured before I stopped.
• But 26 gc::tests::* fail, against a flaky baseline of 1–4 (cargo-test: perry-runtime's suite fails a different number of tests on every run #7365). They are concentrated in debt_pacer, incremental_sweep_reclaim, temp_roots, root_words, triggers — the budgeted/incremental path that scavenge's deferral bypasses. A representative failure asserts a collection count of 1 and gets 0, which is the deferral doing exactly what it is supposed to.

One cause is already understood and fixed in the prototype: force_legacy_gc_pacing() un-capped the trigger by pinning gc_moving_loop_polls_enabled() off, which stops working once the cap is unconditional. The guard now suppresses the cap directly, and that fixed the tests which use it.

But mechanically pinning legacy pacing across incremental_sweep_reclaim fixed only 3 of 10. So the remaining failures are not a pacing-mode mismatch — each needs a judgement call on whether it asserts behaviour that legitimately changed, or catches a real problem. Rushing that is how a real regression gets pinned as "expected".

Suggested shape

1. Land the force_legacy_gc_pacing fix on its own — it is a latent bug today: the guard silently stops working the moment anyone makes the cap unconditional.
2. Triage the ~26 tests individually.
3. Then flip both defaults together, never one alone.
4. Per the kill-policy, the new PERRY_GC_SCAVENGE=0 off-state needs a CI arm; there is none today (grep over .github/workflows returns nothing).

Prototype is on perf/nursery-cap-by-default.

[proggeramlug]

Prototype branch: perf/nursery-cap-by-default (not a PR — see the blocking test work above).