[bompus]

What changed

• Fix: c63e3d7c restores the two-phase resolver pool. perf(resolve): spawn pool workers early and snapshot once after prerequisites #591 had moved the prerequisite (imports/extends/implements) batches to the main-thread kernel so the workers could open one post-prerequisite snapshot; that serialized the whole phase on one core. The pool now opens on the pre-prerequisite snapshot so those refs resolve across the workers, takes a second snapshot once their supertype edges persist, and recycles the workers onto it. It keeps perf(resolve): spawn pool workers early and snapshot once after prerequisites #591's early spawn, single worker-script lookup, loaded message and synthesis-boundary open, and refreshes before synthesis when a sweep had no calls batch.
• Measurement: second 2026-10-31 Node campaign at the fix head against upstream cd04b21a, in docs/benchmarks/fork-vs-upstream-2026-10-31b.md, with the bisect and matched-host A/B.
• Docs: README measured-results rewritten to the fixed run, chart ?v=13, the first 2026-10-31 run and the 2026-10-30 doc keep their dates and now point at the attribution. The MCP memory rows use MiB/1024 (the first run's table and idle list had divided by 1000); the 2026-10-30 3.58 GiB row is labeled 2026-10-27 as its source says.

The regression and its cause

Bisect, same host and corpus state, fresh index per run:

Commit	What it is	supabase	n8n	n8n index CPU
216093df	parent of #590	17.27 / 17.46 s	54.46 / 57.22 s	630–636%
abc462ea	#590 (pool sizing)	17.98 / 18.29 s	55.89 / 58.48 s	626–627%
05077514	#591 (early spawn, one snapshot)	21.84 / 22.51 s	70.34 / 70.70 s	458–462%

#590 alone is inside the pre-#590 spread; its rows ran at 1-minute load 9–11, and both corpora sit above the 25k-refs-per-worker cap. #591 is the regression.

Interleaved A/B at 0d4c48dc vs the fix:

Corpus	Before	After
supabase	21.60 / 21.88 s	17.11 / 17.27 s
n8n	70.72 / 73.17 s	53.58 / 53.91 s

Headline numbers (fixed head vs upstream, spreads in the doc)

Corpus	Index fork/upstream	Sync upstream/fork
gin	1.08× (8% slower, 0.07 s)	1.23×
Alamofire	0.96× (4% faster)	1.39×
pretix	0.89× (11% faster, fork spread 9%)	1.62×
CPython	0.58× (1.71× faster)	2.05×
discourse	0.95× (5% faster)	1.08×
supabase	1.11× (11% slower, 33% more nodes)	2.17×
n8n	0.76× (1.32× faster)	2.69×

MCP n8n: warm explore 240 → 138 ms (1.75×), wave 2.38 → 1.88 s, init 115 → 64 ms, idle 4.46 → 2.19 GiB, busy 3.81 → 3.02 GiB; watcher sync 702 → 766 ms (9% slower).

The fix trades ~0.3 GiB of peak index RSS on n8n for the speed (8.8–8.9 GiB against 8.2–8.5 before; upstream 11.9 GiB), from the second snapshot generation.

Recommendation impact

The 2026-11-02 “continue” recommendation was conditional on this regression being bisected and fixed. It is: index health is restored, and the fork is faster than upstream on five of seven index cells (1.04× to 1.71×) plus all seven sync cells and MCP, with supabase's 11% loss carrying 33% more nodes.

Tests

• Targeted: wal-deferral, resolver-pool-sizing, resolver-pool-engage, kernel-resolve-parity — 51 passed, 1 skipped.
• Full npx vitest run: 11,340 passed, 46 skipped, 4 failed in __tests__/gitignore-pattern-parity.test.ts. Those four also fail at the base commit 0d4c48dc (checked in the primary checkout); they are not touched by this PR.
• check:kernel-rules is red at base too: codegraph-kernel/src/resolve/member_fn_ref.rs:1082 (from feat(kernel): port the remaining upstream resolution drift #587) uses a name-parameterized regex. This PR is TypeScript-only and does not touch it.
• readme-sync: 6/6.
• Campaign admission passed under the host lock (CPU 0.8/4 cores, 98.9 GiB free, memory-full 0/10 ms, I/O-full 9.4/100 ms). Receipt: ~/cg-scratch/bench-2026-10-31d/admission.json.

README rows checked

Upstream merge point (now cd04b21a), measured-results intro, headline chart (?v=13), index table and the graph-size sentence, the new regression-closure paragraph, sync table, MCP table and the idle reconciliation, “What it costs”, and the fork-adds/feature tables (no numbers there moved).

[bompus-ai-review]

Bompus AI Review (Claude Fable 5.1, max reasoning): 0 finding(s).