GC: old-gen fragmentation — scattered survivors pin 105 MB of blocks for a ~1 MB live set

[proggeramlug]

Found while building the 12_large_live_set gc-ratchet probe for #7432 (adaptive tenuring). This is a pre-existing old-gen characteristic that mass promotion now makes routine, not a #7432 regression.

Symptom

A program that promotes a ~100 MB linked structure, then drops all but every 64th node (~1 MB truly live), retains 105.6 MB of arena in-use after an explicit gc(). The full collection itself works — PERRY_GC_DIAG=1 shows freed_bytes=87,679,624 on the final manual collect — but the in-use accounting barely moves:

[gc] blocks: general=50 (49 live), longlived=46 (40 live), freed_bytes=87679624
#gcmetric heap_used_bytes=105588616

49 of 50 blocks still count as live because keeping every 64th node scatters ~218 live objects into every 1 MB block. Block reclaim (arena_reset_empty_blocks / old block reclaim) can only reclaim fully-dead blocks, and the swept holes don't lower block offsets, so heapUsed — and RSS, since the blocks stay mapped — tracks the scattered-survivor high-water forever. V8 compacts old space in this situation; Node reports a proportionally small heap for the same program.

The old-page defrag policy (C4b) did not engage: the traced run reports old_page_candidate_pages: 0 throughout.

Repro

benchmarks/gc_ratchet/probes/12_large_live_set.ts (in #7432), compiled with PERRY_NO_AUTO_OPTIMIZE=1; run with PERRY_GC_DIAG=1. The probe pins today's 105.6 MB deterministically (0% spread), so any improvement must deliberately re-pin.

Why it matters more after #7432

The adaptive tenuring threshold promotes long-lived cohorts on first copy (that is its job — re-copying them was 4.39 GB of waste on tree.ts), so old-gen now receives the scattered-survivor pattern wholesale. The tree.ts residual RSS gap vs PERRY_GC_SCAVENGE=0 (221 MB vs 103 MB) is this plus old-reclaim pacing (#SIBLING), not survivor re-copying.

Directions

• Old-page defrag policy: why zero candidate pages here — thresholds, or the pattern is outside its selection criteria?
• Sparse-block evacuation: blocks whose live fraction is < ~25% are exactly the C4b evacuation case; the survivors are unpinned class instances.
• At minimum, heapUsed could report offsets minus free-listed holes so the metric distinguishes "mapped high-water" from "live".

[proggeramlug]

Sibling RSS follow-up: the tree.ts Pareto cell — see the issue filed just after this one.

[proggeramlug]

Fix up as PR #7443 (stacked on #7432): old-generation hole free list. Probe-12 retained heapUsed drops 105.6 MB -> 59.9 MB; the reclaim pacing arms now subtract reusable hole bytes.