docs: 2026-10-28 seven-corpus re-measure at de5a7b80

README.md

@@ -198,11 +198,11 @@ Call resolution the fork adds:
 
 ### Measured results
 
-Upstream `main` at `69af19ac` against the fork at `645abbb8`, each run on Node.js 24.22.1 and on Bun 1.4.4. Index and sync figures were re-measured 2026-10-27 on the same 16-vCPU WSL2 host after the ambient-`declare module` scan fix and the resolve-pipeline changes below; MCP, command-line and explore figures are carried from the earlier 2026-10-27 run at `bc27fdb4` (unchanged by resolution-path changes). Seven corpora, arms in mirrored order; index figures are the median of two runs and sync figures the median of four. A comparison counts as a change only when the gap exceeds that comparison's own run-to-run spread — the larger of the two arms' (max − min) / median — and cells that cannot show that are marked inconclusive or no clear change rather than a fixed band. Upstream refuses to start on Bun, which reports itself as Node 26, because upstream blocks Node 25 and newer; it ran here with `CODEGRAPH_ALLOW_UNSAFE_NODE=1`. The host shared other work during the run, so spread-based cells carry more noise than a quiet host would; the numbers, the method, the graph sizes and the disclosures are in [`docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md`](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md). The scripts that produced them are in [`scripts/benchmarks/runtime/`](scripts/benchmarks/runtime/).
+Upstream `main` at `69af19ac` against the fork at `de5a7b80`, each run on Node.js 24.22.1 and on Bun 1.4.4. Index and sync figures were re-measured 2026-10-28 on the same 16-vCPU WSL2 host after the ambient-`declare module` scan fix and the resolve-pipeline changes below; MCP, command-line and explore figures are carried from the earlier 2026-10-27 run at `bc27fdb4` (unchanged by resolution-path changes). Seven corpora, arms in mirrored order; index figures are the median of two runs and sync figures the median of four. A comparison counts as a change only when the gap exceeds that comparison's own run-to-run spread — the larger of the two arms' (max − min) / median — and cells that cannot show that are marked inconclusive or no clear change rather than a fixed band. Upstream refuses to start on Bun, which reports itself as Node 26, because upstream blocks Node 25 and newer; it ran here with `CODEGRAPH_ALLOW_UNSAFE_NODE=1`. The host shared other work during the run, so spread-based cells carry more noise than a quiet host would; the numbers, the method, the graph sizes and the disclosures are in [`docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md`](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md). The scripts that produced them are in [`scripts/benchmarks/runtime/`](scripts/benchmarks/runtime/).
 
-![Fork against upstream CodeGraph on n8n and the command line: the fork indexes n8n 1.5 times faster, syncs an edited file 1.3 times faster, answers a warm question 1.8 times faster and uses 33% less memory while indexing. The fork is better in 57 of 76 comparisons, shows no clear change in 10 and is worse in 9.](assets/fork-vs-upstream-performance.svg?v=8)
+![Fork against upstream CodeGraph on n8n and the command line: the fork indexes n8n 1.7 times faster, syncs an edited file 3.5 times faster, answers a warm question 1.8 times faster and uses 30% less memory while indexing. The fork is better in 57 of 76 comparisons, shows no clear change in 6 and is worse in 13.](assets/fork-vs-upstream-performance.svg?v=9)
 
-The fork's one-file sync is faster than upstream's on six of the seven corpora on Node — 4% (pretix) to 56% (CPython), with supabase the one no-clear-change — and on Bun it now wins every clear comparison, pretix included (0.68 s against ~1.13 s after the retry fix below). Full indexing: the fork wins on n8n (1.5×), CPython (2.2×), discourse (1.4×), pretix-Node (8.14 s against 9.04 s) and the Bun arm of gin, Alamofire, discourse, supabase and n8n; gin-Node, Alamofire-Node and supabase-Node tie inside spread; pretix-Bun is within ~2% of upstream (the last residual). The speed costs some memory: indexing peaks higher on pretix (roughly +22% Node / +26% Bun after the worker-cache halving below) and supabase-Bun — the 12 in-process worker heaps are the disclosed trade for the index-time wins — while of the other five corpora on Node, four index at 4–32% less and Alamofire runs 1% higher (inside its spread). Cells where the fork still trails: pretix-Bun index (~+2%), the pretix and supabase-Bun index-memory cells, and a ~0% gin file-watcher cell inside noise. Three earlier losses — pretix-Bun sync, n8n busy-MCP memory on Node and the gin concurrent-explore wave — are wins in this tree; the post-merge measurements are in the benchmark doc's [follow-up section](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#follow-up-merges). The benchmark doc has [every result](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#every-result-at-a-glance) in one heatmap.
+The fork's one-file sync is faster than upstream's on six of the seven corpora on Node — 45% (pretix) to 72% (n8n), with discourse the one loss (+47%) — and on Bun it wins six of seven the same way, discourse again the loss (+31%). Full indexing: the fork wins on n8n (1.7×), CPython (2.1×), discourse (1.6×), pretix-Node (7.48 s against 9.69 s), supabase-Node (15.6 s against 19.6 s) and Alamofire-Node (1.19 s against 1.42 s), and on the Bun arm of Alamofire, pretix (7.15 s against 7.20 s — a tie inside spread), CPython, discourse, supabase and n8n; gin index is the remaining loss (+6% Node / +18% Bun — ~0.05–0.14 s of fixed fork overhead on a 119-file index). The speed costs some memory: indexing peaks higher on pretix (+36% Node / +34% Bun) and supabase (+14% Node / +13% Bun) — the 12 in-process worker heaps are the disclosed trade for the index-time wins — while the fork now peaks lower than upstream on n8n (−30%), CPython (−9%), gin (−33%) and discourse-Bun (−28%). Cells where the fork still trails: gin index on both runtimes, discourse sync on both runtimes, and the pretix and supabase index-memory cells. The post-merge measurements are in the benchmark doc's [follow-up section](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#follow-up-merges). The benchmark doc has [every result](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#every-result-at-a-glance) in one heatmap.
 
 The fork starts its resolver pool at or above 150,000 pending references. Below that, it can start the pool part-way through if at least 20,000 unresolved references remain and the estimated main-thread work is at least 0.4 s. An explicit `CODEGRAPH_PARALLEL_RESOLVE_MIN` replaces both gates. Resolution proceeds in pages of 40,000 references (`CODEGRAPH_RESOLVE_BATCH` overrides), across up to 12 workers (`CODEGRAPH_RESOLVE_WORKERS` sets the count, `0` disables the pool; `CODEGRAPH_NO_PARALLEL_RESOLVE=1` does the same). Against a pre-change baseline build, five tested corpora indexed 11-48% faster at 56-132% more peak memory with identical graphs ([measurements](docs/benchmarks/adaptive-pool-engage-2026-10-15.md)).
 
@@ -212,29 +212,29 @@ A [2026-10-15 mid-size comparison](docs/benchmarks/midsize-sweep-2026-10-15.md)
 
 | Corpus | Upstream, Node | Fork, Node | Fork, Bun |
 |---|---|---|---|
-| gin (Go, 119 files) | 0.96 s, 558 MiB | 1.02 s, 399 MiB | 0.88 s, 308 MiB |
-| Alamofire (Swift, 129 files) | 1.35 s, 879 MiB | 1.41 s, 891 MiB | 0.94 s, 881 MiB |
-| pretix (Python + JS, 1,472 files) | 9.04 s, 3.29 GiB | 8.14 s, 4.17 GiB | 7.49 s, 3.79 GiB |
-| CPython (C + Python, 3,708 files) | 55.0 s, 7.15 GiB | 25.2 s, 5.90 GiB | 25.1 s, 5.39 GiB |
-| discourse (Ruby + JS, 20,277 files) | 19.6 s, 4.07 GiB | 13.7 s, 3.92 GiB | 12.1 s, 2.97 GiB |
-| supabase (React + Next.js + TS, 10,716 files) | 18.2 s, 4.86 GiB | 15.5 s, 5.45 GiB | 14.6 s, 4.73 GiB |
-| n8n (Vue + TS, 24,434 files) | 74.9 s, 12.39 GiB | 49.5 s, 8.26 GiB | 45.2 s, 8.28 GiB |
+| gin (Go, 119 files) | 0.96 s, 567 MiB | 1.02 s, 381 MiB | 0.91 s, 310 MiB |
+| Alamofire (Swift, 129 files) | 1.42 s, 817 MiB | 1.19 s, 1.05 GiB | 1.01 s, 887 MiB |
+| pretix (Python + JS, 1,472 files) | 9.69 s, 2.92 GiB | 7.48 s, 3.97 GiB | 7.15 s, 3.85 GiB |
+| CPython (C + Python, 3,708 files) | 53.7 s, 6.21 GiB | 25.5 s, 5.64 GiB | 25.6 s, 5.41 GiB |
+| discourse (Ruby + JS, 20,277 files) | 21.7 s, 3.59 GiB | 13.2 s, 3.58 GiB | 12.4 s, 2.98 GiB |
+| supabase (React + Next.js + TS, 10,716 files) | 19.6 s, 4.23 GiB | 15.6 s, 4.83 GiB | 14.5 s, 4.63 GiB |
+| n8n (Vue + TS, 24,434 files) | 83.0 s, 11.00 GiB | 49.4 s, 7.65 GiB | 46.6 s, 8.43 GiB |
 
-The fork's index was slower than upstream's on five of the seven corpora in the first pass of this run — the supabase and n8n regressions traced to upstream's ambient-`declare module` lookup (merged in fork #548): the Rust port scanned every declaration file per out-of-repo import. A per-file memo ([#554](https://github.com/bompus/codegraph/pull/554)) ended the reparse-every-reference quadratic, and the follow-up above lands the rest: a port of upstream's text tokenizer plus a `specifier → files` index, so each import checks only candidate declaration files instead of every `.d.ts`. The re-measured numbers in this table are the result: supabase went from 41.5 s to 15.5 s on Node and n8n from 178 s to 49.5 s — n8n now beats upstream 1.5× and CPython 2.2×. The remaining index losses from that run — pretix +33% and Alamofire +18% on Node, and the Bun arm of supabase and n8n — traced not to the ambient scan but to the resolve loop's batch boundaries: every 5,000-ref page drained the worker pipeline before the next fanned out. This run stacks three fixes for that — 40,000-ref pages with a 12-worker pool cap (same batch-boundary semantics; pool workers resolve against a fixed snapshot), then bounded work-stealing dispatch with earlier batch fan-out and non-blocking worker recycling, then a memoized near-duplicate pass that stopped re-tokenizing bodies it had just signed — and those losses are gone: pretix-Node now wins (8.14 s against 9.04 s), the Bun arms of supabase and n8n win, and pretix-Bun is within ~2% of upstream (the last residual). Later in the same stack, the deferred-decode path stopped turning off on framework projects: a framework `extract()` hook no longer forces every file onto decoded-object extraction — its output rides the kernel-buffer transport and merges post-decode ([#563](https://github.com/bompus/codegraph/pull/563); pretix went from 0 to 1,416 kernel-path files, n8n-Node index RSS −10%). The benchmark doc has the [full accounting](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#the-supabase-resolution-regression). The trade is memory: indexing peaks higher on pretix and supabase now, lower on the other five, and its one-file sync is faster on six of seven (below).
+The fork's index was slower than upstream's on five of the seven corpora in the first pass of this run — the supabase and n8n regressions traced to upstream's ambient-`declare module` lookup (merged in fork #548): the Rust port scanned every declaration file per out-of-repo import. A per-file memo ([#554](https://github.com/bompus/codegraph/pull/554)) ended the reparse-every-reference quadratic, and the follow-up above lands the rest: a port of upstream's text tokenizer plus a `specifier → files` index, so each import checks only candidate declaration files instead of every `.d.ts`. The re-measured numbers in this table are the result: supabase went from 41.5 s to 15.6 s on Node and n8n from 178 s to 49.4 s — n8n now beats upstream 1.7× and CPython 2.1×. The remaining index losses from that run — pretix +33% and Alamofire +18% on Node, and the Bun arm of supabase and n8n — traced not to the ambient scan but to the resolve loop's batch boundaries: every 5,000-ref page drained the worker pipeline before the next fanned out. This run stacks three fixes for that — 40,000-ref pages with a 12-worker pool cap (same batch-boundary semantics; pool workers resolve against a fixed snapshot), then bounded work-stealing dispatch with earlier batch fan-out and non-blocking worker recycling, then a memoized near-duplicate pass that stopped re-tokenizing bodies it had just signed — and those losses are gone: pretix-Node wins (7.48 s against 9.69 s), the Bun arms of supabase and n8n win, and pretix-Bun is a tie (7.15 s against 7.20 s). Later in the same stack, the deferred-decode path stopped turning off on framework projects: a framework `extract()` hook no longer forces every file onto decoded-object extraction — its output rides the kernel-buffer transport and merges post-decode ([#563](https://github.com/bompus/codegraph/pull/563); pretix went from 0 to 1,416 kernel-path files, n8n-Node index RSS −10%). The benchmark doc has the [full accounting](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-27.md#the-supabase-resolution-regression). The trade is memory: indexing peaks higher on pretix and supabase now, lower on the other five, and its one-file sync is faster on six of seven (below).
 
-Upstream at `6560052` took 2.0 to 2.7 times as long to index Python as at `290e03f` (pretix 9.2 s to 18.8 s, CPython 35.6 s to 95.3 s; [a same-host check](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-03.md#since-2026-09-28) put the cause in upstream's code). At `69af19ac` upstream's CPython index is slower than at `986de95c` (55.0 s against 40.2 s on this host) while the fork's keeps improving, so the fork's CPython lead is 2.2× this run.
+Upstream at `6560052` took 2.0 to 2.7 times as long to index Python as at `290e03f` (pretix 9.2 s to 18.8 s, CPython 35.6 s to 95.3 s; [a same-host check](docs/benchmarks/fork-vs-upstream-node-bun-2026-10-03.md#since-2026-09-28) put the cause in upstream's code). At `69af19ac` upstream's CPython index is slower than at `986de95c` (53.7 s against 40.2 s on this host) while the fork's keeps improving, so the fork's CPython lead is 2.1× this run.
 
 **One-file sync** (edit one file, `codegraph sync`):
 
 | Corpus | Upstream, Node | Fork, Node | Fork, Bun |
 |---|---|---|---|
-| gin | 0.47 s | 0.40 s | 0.29 s |
-| Alamofire | 0.48 s | 0.29 s | 0.18 s |
-| pretix | 1.62 s | 1.55 s | 0.68 s |
-| CPython | 5.35 s | 2.38 s | 1.66 s |
-| discourse | 4.78 s | 2.37 s | 1.77 s |
-| supabase | 4.58 s | 2.76 s | 1.95 s |
-| n8n | 11.6 s | 9.07 s | 8.10 s |
+| gin | 0.34 s | 0.27 s | 0.19 s |
+| Alamofire | 0.42 s | 0.27 s | 0.17 s |
+| pretix | 1.56 s | 0.86 s | 0.60 s |
+| CPython | 6.01 s | 2.65 s | 1.72 s |
+| discourse | 1.19 s | 1.75 s | 1.20 s |
+| supabase | 8.17 s | 3.05 s | 1.94 s |
+| n8n | 11.6 s | 3.29 s | 2.45 s |
 
 Upstream's supabase sync works in this run — the `FOREIGN KEY constraint failed` seen in every prior run ([colbymchenry/codegraph#2565](https://github.com/colbymchenry/codegraph/issues/2565)) was fixed upstream on 2026-10-19. Its two passes differed widely on Node (1.05 s and 8.1 s); the 4.58 s figure is the median of four, and the spread means that comparison counts as no clear change rather than a fork win.