Perf: reclaim GC slab slots with per-chunk free state

Parent: #97 Related: #62, #101, zig-utils/zig-gc#17

Evidence

After bounded tail trimming (#101), direct and shared object_churn profiles still spent most nursery time in sweep and GcCellBacking.freeFn. The rejected zig-gc#17 experiment proved that batching allocator dispatch/size-class locks was neutral at the 8x100 publication load: per-cell ownership classification and writing/linking one intrusive FreeNode remained the real work.

Retained design

Reusable slots now live in fixed-size per-chunk bitmaps, with a hot reusable-chunk hint per size class. Sweep no longer writes allocator links into dead cells. Tail trimming counts and removes bitmap state only for discarded chunks.

Implementation: 4e3de6f5 Publication: 24656d74 Pre-existing amplify seed 107 timeout: #103 Pre-existing midgc CI case: #96

Acceptance criteria

• Reuse freed slots without one intrusive linked-list write per dead cell.
• Preserve partially live chunks, mixed size classes, stale-header rejection, ownership indexes, exact stats, and bounded tail trimming.
• Prove wholly empty and fragmented chunk reuse, full/minor collections, OOM recovery, teardown, and parallel allocation/sweep safety.
• Exact-parent direct and 1/2/4/8-lane shared object-churn A/B retains exact checksums and materially improves the publication load.
• Pass focused allocator/GC/no-GIL units, full units once, and relevant TSan/threadfuzz gates; deterministic unrelated failures reproduce on the exact parent and are tracked as #96/#103.
• Refresh the full report, raw TSV, README, and benchmark guide after a retained change.

Implementation is now on main as 4e3de6f5 (perf(gc): track reusable slots outside cells (#102)).

What changed:

• replaced intrusive freed-cell FreeNode writes with fixed-size per-chunk bitmaps
• added a hot reusable-chunk hint while keeping exact per-bucket locking
• made tail trimming count/drop bitmap state only for discarded chunks
• preserved maintained capacity/issued/live/free/fresh/reused/freed counters
• added fragmented reuse coverage proving the backing does not write allocator metadata into freed cell payloads

Exact-parent A/B against untouched e4b72407 (checksums exact):

• direct object_churn, 100 jobs, 21 samples: 190.549 -> 177.473 ms (7.37% faster)
• shared 8x10, 21 samples: 511.461 -> 513.864 ms (-0.47%, neutral)
• publication-load shared 8x100, 6 samples: 8736.301 -> 8485.471 ms (2.96% faster)

Validation:

• 21/21 focused allocator units
• 763/763 full units
• 21/21 focused allocator units under TSan
• amplify TSan fuzzing isolated deterministic seed 107 timeout; the exact parent reproduces identically, so tracked separately as #103 and not caused by this change

The full 1,540-sample JSC comparison and README/docs refresh are next; keeping this issue open until publication and CI are complete.

Publication and the remaining exact-parent lane sweep are complete.

Exact-parent shared object_churn A/B at the 100-job publication load (two alternating rounds, six samples per binary/lane, exact checksums):

• 1 lane: 375.592 -> 371.525 ms (1.09% faster)
• 2 lanes: 753.025 -> 734.590 ms (2.51% faster)
• 4 lanes: 2012.021 -> 1765.678 ms (13.95% faster)
• 8 lanes: 8736.301 -> 8485.471 ms (2.96% faster)

Published in 24656d74 from measured clean commit 4e3de6f5:

• full report: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.md
• 1,540 raw samples: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.tsv
• README and benchmark guide refreshed

New full-matrix headline:

• direct: 9/10 wins; 2.56x geometric-mean throughput lead over JSC
• warmed 8 lanes: 2.54x lead; scaling 5.18x vs JSC 5.32x
• cold 8 lanes: 2.58x lead; scaling 5.29x vs JSC 5.25x
• shared 8-lane geometric-mean scaling: 4.09x
• object_churn: 176.398 ms direct and 9203.183 ms shared 8-lane

The full run captured battery power at 93% and discharging, which is disclosed in the report/README. Cross-publication changes include host variation; the interleaved exact-parent A/B above is the causal evidence.

Completed on main in 4e3de6f5 with publication 24656d74. All acceptance criteria are checked above; exact-parent direct and 1/2/4/8 shared A/B retained exact checksums and improved every measured point. Follow-up fuzz/lifecycle failures that reproduce on the parent remain tracked independently in #96 and #103.