refactor: extract SlotPacker (R1 live-slot packing) — extraction PR 2

[Xof]

Task 3 of the transaction.rs extraction (after #76). Extracts SlotPacker — an owned state unit for the R1 live-slot / insert-cursor model — out of TransactionManager. Pure behavior-preserving refactor; the existing suite is the oracle.

What changed

• SlotPacker (in transaction/packing.rs) now owns the three fields committed_live_slots / current_live_slots / insert_cursor and the packing logic. Narrow interface: new / from_committed, insert(cache, alloc, packing_enabled, value), release -> Option<u64>, begin/commit/rollback, snapshot/restore/clear_cursor, plus read accessors. No code outside packing.rs touches those fields directly (surface-checked).
• Borrow crux: SlotPacker::insert holds &mut self.packer while needing a fresh data page, whose allocation touches the disjoint freemap field cluster. The freemap take/cow_alloc/put dance is hoisted to the caller (insert_into_data_page) around the alloc closure — the same disjoint-field-borrow pattern ht_insert already uses. alloc receives cache as a param, captures only the freemap field refs.
• Three plan-interface gaps resolved from the actual code: insert takes a packing_enabled bool (SlotPacker stays ignorant of savepoints); release returns Option<u64> so the caller owns the txn_freed_pages push; from_committed seeds the open-time scan (vs empty new).
• The now-dead TransactionManager::allocate_data_page (its sole caller inlined the dance) was removed; its cross-file comment references were reconciled to point at cow_alloc (second commit) — including three that were already loose pre-refactor (persist_freemap allocates freemap pages via structural_extend, never data pages).

Verification

• cargo test → 578 passing, 0 failed (273 chisel-lib) — identical to the pre-refactor baseline
• cargo clippy --workspace --all-targets -- -D warnings clean; cargo fmt --check clean
• pytest → 119 passing (binding untouched)
• Surface check: the three packer fields appear outside packing.rs only as scan locals, the Savepoint struct's own snapshot fields, and comments.

Reviewed

Call-site rewires (recovery/lifecycle/savepoints/stats/tests) independently verified semantically faithful; the freemap-hoist error-path ordering and all-three-lifecycle-clear-cursor invariants confirmed preserved.

Next: Task 4 (FreemapRecycle — the hardest), one PR off updated main.

[github-actions]

🚦 Bench results: PR vs main

⚠️ 3 regression(s) detected across 2 scenario/mode pair(s)

Scenario	Mode	Δ throughput	Worst Δ
mutation-log	sqlite-strict	+0.0%	p99 +11.7% ⚠️
ycsb-a	chisel-mem	-7.9%	p50 +8.3% ⚠️
document-store	chisel-mem	+1.3%	—
document-store	chisel-strict	+5.5%	—
document-store	redb-strict	+3.1%	—
document-store	sqlite-strict	+0.4%	—
mutation-log	chisel-mem	-4.6%	—
mutation-log	chisel-strict	-0.1%	—
mutation-log	redb-strict	-0.2%	—
ycsb-a	chisel-strict	+0.1%	—
ycsb-a	redb-strict	+0.1%	—
ycsb-a	sqlite-strict	-0.2%	—
ycsb-b	chisel-mem	+0.2%	—
ycsb-b	chisel-strict	-2.1%	—
ycsb-b	redb-strict	+0.2%	—
ycsb-b	sqlite-strict	+0.6%	—

Per-scenario detail (4 metrics × cells)

document-store

Mode	Throughput	p50	p95	p99
chisel-mem	18530 ops/s → 18777 ops/s (+1.3%)	14.9 µs → 14.8 µs (-0.8%)	137.9 µs → 136.2 µs (-1.2%)	671.6 µs → 655.0 µs (-2.5%)
chisel-strict	2938 ops/s → 3099 ops/s (+5.5%)	24.7 µs → 22.8 µs (-7.6%)	908.4 µs → 819.5 µs (-9.8%)	3.14 ms → 2.87 ms (-8.8%)
redb-strict	4009 ops/s → 4132 ops/s (+3.1%)	12.5 µs → 12.0 µs (-3.7%)	511.3 µs → 439.2 µs (-14.1%)	1.39 ms → 1.26 ms (-9.5%)
sqlite-strict	5131 ops/s → 5152 ops/s (+0.4%)	21.2 µs → 20.9 µs (-1.4%)	363.0 µs → 345.2 µs (-4.9%)	1.33 ms → 1.29 ms (-3.2%)

mutation-log

Mode	Throughput	p50	p95	p99
chisel-mem	51669 ops/s → 49293 ops/s (-4.6%)	23.0 µs → 23.1 µs (+0.3%)	28.7 µs → 29.6 µs (+3.2%)	35.8 µs → 39.2 µs (+9.2%)
chisel-strict	1773 ops/s → 1771 ops/s (-0.1%)	296.7 µs → 296.0 µs (-0.2%)	659.4 µs → 649.1 µs (-1.6%)	17.96 ms → 18.29 ms (+1.8%)
redb-strict	1806 ops/s → 1802 ops/s (-0.2%)	146.5 µs → 147.8 µs (+0.9%)	231.8 µs → 243.4 µs (+5.0%)	34.40 ms → 33.57 ms (-2.4%)
sqlite-strict	5031 ops/s → 5032 ops/s (+0.0%)	97.2 µs → 96.9 µs (-0.4%)	267.4 µs → 270.2 µs (+1.1%)	370.1 µs → 413.3 µs (+11.7%) ⚠️

ycsb-a

Mode	Throughput	p50	p95	p99
chisel-mem	37710 ops/s → 34741 ops/s (-7.9%) ⚠️	41.0 µs → 44.4 µs (+8.3%) ⚠️	60.1 µs → 64.5 µs (+7.2%)	77.1 µs → 83.8 µs (+8.7%)
chisel-strict	2262 ops/s → 2263 ops/s (+0.1%)	255.6 µs → 256.2 µs (+0.2%)	743.7 µs → 724.2 µs (-2.6%)	1.58 ms → 1.71 ms (+8.2%)
redb-strict	2681 ops/s → 2683 ops/s (+0.1%)	116.3 µs → 116.0 µs (-0.3%)	202.2 µs → 196.6 µs (-2.8%)	389.7 µs → 326.9 µs (-16.1%)
sqlite-strict	132941 ops/s → 132625 ops/s (-0.2%)	7.6 µs → 7.7 µs (+0.8%)	10.1 µs → 10.1 µs (+0.3%)	11.8 µs → 12.2 µs (+2.9%)

ycsb-b

Mode	Throughput	p50	p95	p99
chisel-mem	162521 ops/s → 162861 ops/s (+0.2%)	4.7 µs → 4.7 µs (-1.1%)	42.8 µs → 43.8 µs (+2.2%)	49.8 µs → 49.4 µs (-0.7%)
chisel-strict	21842 ops/s → 21381 ops/s (-2.1%)	7.5 µs → 7.1 µs (-6.4%)	263.8 µs → 262.2 µs (-0.6%)	669.5 µs → 706.3 µs (+5.5%)
redb-strict	27223 ops/s → 27269 ops/s (+0.2%)	3.1 µs → 3.1 µs (-0.3%)	119.5 µs → 116.2 µs (-2.8%)	191.0 µs → 178.5 µs (-6.6%)
sqlite-strict	161560 ops/s → 162600 ops/s (+0.6%)	6.3 µs → 6.3 µs (-0.9%)	8.4 µs → 8.3 µs (-1.1%)	10.6 µs → 10.6 µs (+0.4%)

_{Generated by chisel-bench-diff at 2026-06-23T03:12:33Z. Compares PR HEAD against main. Never blocks merge — signal, not gate. Thresholds: throughput 5%, p50 5%, p95 10%, p99 10%.}