perf(signals): mapArray stages only the changed window and commits in place

.changeset/map-array-subsequence-fast-paths.md

@@ -0,0 +1,5 @@
+---
+"@solidjs/signals": patch
+---
+
+`mapArray` commits a structural change without rebuilding full-length copies of its row bookkeeping: a pass stages only the changed window, lands it into the live arrays in place with the unchanged suffix shifted by a loop, and rebuilds only the mapped array that signals the change downstream. A pass that only removes rows is matched with two pointers and no key map, and the key-chain array is sized to the window rather than to the edit's position. One-item inserts and removals on long keyed lists, every-nth removals and sparse filters no longer allocate several list-sized arrays per edit; the strong-abort contract for a mapper that suspends is unchanged, kept by ordering (verify, create, then mutate) instead of by copying.

documentation/benchmarking-strategy.md

@@ -392,26 +392,51 @@ campaign that built this and what it found is the
 [Store / List Delivery Lane](./performance-experiments.md#store--list-delivery-lane-2026-08-16--2026-09-18-the-octane-board)
 in the journal. The mechanics:
 
-**Where.** Fork `ryansolid/octane`, branch `solid-rc8-fixtures-rebase`
-(upstream `main` + our Solid fixture commits; rebase it when upstream
-moves — the Solid fixtures are ours alone, upstream still pins
-2.0-beta.20). Local checkout `~/Development/octane-fork`. A second
-checkout, `~/Development/octane`, holds the pre-September experiment
-scaffolding and the older results directories.
-
-**Linking `next`.** Keep a worktree of this repo on `origin/next`
-(`~/Development/solid-next-bench`), `pnpm install`, `pnpm build` in
-`packages/compiler` (native binary, needs cargo), then
+**Where.** Fork `ryansolid/octane`, branch `bench/next-linked-oct15`
+(upstream `octanejs/octane` `main` at 0.13.8, `91dae8d73`, plus our seven
+Solid fixture commits — the shape passes, `board-compare.mjs`, the
+`uibench/solid-reconcile` target and its manifest line). Local checkout
+`~/Development/octane-fork`. Upstream pins the published rc (rc.14 on
+2026-10-15) and never carries our shapes, so **rebase before every board**:
+`git checkout -b bench/next-linked-<date> && git fetch origin main && git
+rebase origin/main`. Upstream edits the same fixtures' `package.json` /
+`vite.config.js` identically to our pin commit, so the only conflicts are
+`pnpm-workspace.yaml` and `pnpm-lock.yaml` — take upstream's side for both
+(`git checkout --ours`), the pin commit drops out as empty, and the lockfile
+is regenerated by the install below. Older branches are kept as they were:
+`bench/next-linked` and `solid-rc8-fixtures-rebase` are the 0.2.13-era
+boards (Sep 9 / Sep 17). A second checkout, `~/Development/octane`, holds
+the pre-September experiment scaffolding and the older results directories.
+
+**Linking `next`.** Keep a worktree of this repo at the `next` commit under
+test (`~/Development/solid-next-bench`; `git worktree add -b
+bench/next-<date> ~/Development/solid-next-bench <sha>` — a worktree cannot
+check out `next` itself while the main checkout is on it, hence the branch),
+`pnpm install --frozen-lockfile`, `pnpm --filter @solidjs/compiler build`
+(native binary, needs cargo, ~30 s), then
 `pnpm turbo run build --filter=@solidjs/signals --filter=solid-js
---filter=@solidjs/web --filter=@solidjs/babel-plugin --filter=@solidjs/compiler`.
+--filter=@solidjs/web --filter=@solidjs/babel-plugin` (~25 s).
 In the fork's `pnpm-workspace.yaml` `overrides:` block, point `solid-js`,
 `@solidjs/web`, `@solidjs/signals`, `@solidjs/compiler` and
 `@solidjs/babel-plugin` at `link:../solid-next-bench/packages/<name>`,
-then `pnpm install --no-frozen-lockfile`. `@solidjs/vite-plugin` stays
-npm and picks the linked compiler up through the override — verify with
+then `pnpm install --no-frozen-lockfile` (the fork is on pnpm 12; ~40 s
+with a warm store). `@solidjs/vite-plugin` stays npm (3.0.0-next.49 drives
+the compiler built from `next` at rc.15 without complaint) and picks the
+linked compiler up through the override — verify with
 `readlink -f node_modules/.pnpm/@solidjs+vite-plugin*/node_modules/@solidjs/compiler`.
 **Keep the override and the lockfile it produces uncommitted**; the
 fixture branch must install from the published rc for anyone else.
+The 2026-10-15 session found the worktree deleted and the overrides
+dangling — the board was unrunnable until it was recreated; check
+`ls ~/Development/solid-next-bench` before trusting the link.
+
+**Solid-vs-Solid A/B.** A second worktree at the reference commit
+(`~/Development/solid-sep17-bench` at `dd908fc12`, the Sep 17 board's
+`next`), built the same way. Flip the five overrides to it, reinstall,
+run the suites into one results dir, flip back, run again into another;
+the runner rebuilds the fixtures itself. Same machine, same Octane, same
+day — this is the only read on "did `next` regress" that does not depend
+on the harness staying still between boards (it did not: see gotchas).
 
 **Running.** From the fork root:
 
@@ -441,6 +466,21 @@ for the two-sided geomean board (timed ops only). For a single suite,
 
 **Gotchas, each learned the hard way:**
 
+- **The machine must be idle for the whole run, and nobody but the
+  runner can tell.** `ps -Ao pcpu,etime,comm -r | head` before launching
+  and again when it finishes. The 2026-10-15 board's last two suites ran
+  beside a game and a Chrome renderer each holding more than half a core:
+  every target's uibench numbers came out 4–5× slower than the quick pass
+  an hour earlier and the Solid/Octane ratio there read 1.90× instead of
+  1.63× — load does not cancel in the ratio. The reference frameworks'
+  absolute drift against the previous board (React, Inferno, Svelte,
+  Preact; pins unchanged) is the tell: 0.97–1.05 on a clean suite, anything
+  else means rerun that suite.
+- Upstream ≥ 0.13 batches the sub-millisecond js-framework ops (`select`,
+  `swap`, `select_lots`: `BATCHED_OPS` in `js-framework/run.mjs`) inside each
+  sample. Those ops now read 10–40× smaller than on the 0.2.13 boards for
+  every framework; compare them only within a run. uibench always
+  repeated its small cases, so its numbers are comparable across boards.
 - `PLAYWRIGHT_BROWSERS_PATH` is set in the agent environment to a sandbox
   cache; unset it or Chromium "doesn't exist".
 - Long runs and preview servers must be tool-managed background tasks
@@ -467,13 +507,22 @@ for the two-sided geomean board (timed ops only). For a single suite,
   until the fixture has been read.
 
 **Profiling a single op.** Build the fixture unminified
-(`npx vite build --minify false --outDir dist-prof`), serve it, and drive
+(`npx vite build --minify false --outDir dist-prof`), serve it
+(`npx vite preview --outDir dist-prof --port 5347 --strictPort`), and drive
 the op from a Playwright script with CDP `Profiler` (self-time by
 function) and `HeapProfiler.startSampling` (allocation by function);
-interleave the A and B pages in one browser for paired comparison. The
-2026-09-18 todomvc scripts (`todo-prof.mjs`, `todo-var.mjs`,
-`todo-alloc.mjs`, `todo-ab.mjs`) are the template; they lived in `/tmp`
-and should be re-created next to the harness if this becomes routine.
+interleave the A and B pages in one browser for paired comparison. For
+uibench this is `benchmarks/uibench/tools/prof.mjs` in the fork (untracked;
+run from `benchmarks/uibench` so it resolves the suite's Playwright):
+`node tools/prof.mjs --url=http://localhost:5347/ --case='tree/[500]/[removeLast(1)]'
+--reps=200 --mode=both`. It follows `run.mjs`'s protocol (prepare → settle →
+gc → run) and brackets each `__run` with two sentinel functions that spin
+0.3 ms so they are always sampled; only the samples between them count, so
+the `__prepare` of each rep — a full render for the removeAll cases — stays
+out of the numbers. Allocation sampling is started and stopped around each
+`__run` over CDP for the same reason. The 2026-09-18 todomvc scripts
+(`todo-prof.mjs`, `todo-var.mjs`, `todo-alloc.mjs`, `todo-ab.mjs`) were the
+template; they lived in `/tmp` and are gone.
 
 ## Cross-references