perf: sort interactions at TanStack-beating latency, end to end (#457)

[blove]

Closes #457. Three design cycles on one branch, each specced and measured (specs and plans under docs/superpowers/).

What this delivers, measured (S2 bench, like-for-like, same-run TanStack control)

metric	before	after
50k sort status	partial ×3 (never settles)	completed ×3
50k sort interaction latency	unreportable (~390–409ms once completing)	14.8–16.7ms (TanStack same-run: 35–49ms)
3k sort latency	50–59ms	8.4–23ms
3k filter-metadata (collateral)	42–58ms	15.8ms
50k mount first paint	~450ms blank	painted in the activation pass

How, per cycle

Cycle 1 — synchronous sort fast path (row-model). A conservative query-delta classifier routes a sort-only setQuery on ungrouped data to a synchronous rebuild: no cooperative slices (browser-measured: scheduler hops cost frames), O(n) balanced bulk tree construction from sorted entries. Error/supersede semantics mirror the cooperative path observable-for-observable.

Cycle 2 — sort keys move to the plan; the journal learns "order-only". CompiledRowMetadata lost sortKeys; each compiled plan owns a WeakMap sort-key store, so a sort-only change carries records and the rows HAMT by identity (rebuilding them was 86% of the transition cost). The change journal gained reset reason "reorder" (fail-closed for unaware consumers), layout-core gained a synchronous RowHeightIndex.reorder() over existing measurements, and the row-layout controller takes a permutation path instead of re-ingesting 50k rows.

Cycle 3 — decorated entries + bulk mount. Tree entries carry their resolved sort keys ({record, keys}) so comparisons are property reads — this recovered the grouped gate regression that per-comparison WeakMap resolution had introduced, and improved every comparator-hot path (hence the filter gains). Replacements over unmeasured bases (true mounts) build synchronously in one O(n) pass, eliminating the blank-mount window and the sort-during-ingest race; a reorder arriving mid-replacement now composes at finish instead of forcing a full re-ingest.

Verification

• Every task TDD'd and mutation-hardened (each load-bearing assertion demonstrated to fail under a seeded defect); equivalence oracles against cold-built models throughout.
• Suites: row-model 412, layout-core 118, renderer-dom 144, react 1246, grid-core 125 (incl. a fail-closed reorder-reset pin), website 557 — all green; pnpm api drift-free (public surface change: the reset-reason union gains "reorder", core+react reports regenerated).
• Browser traces confirm the benched 50k sort takes the reorder path (reorderFallbackCount 0, no re-ingest frames) and the mount publishes in the activation pass.
• Full measurement records: docs/superpowers/specs/*2026-08-1{7,8}* reference the scratchpad protocol; cycle-by-cycle numbers in the PR conversation on request.

Merge checklist

• CI green
• Grouped gate quiet-machine certification (pnpm bench:row-model:gate ×3 ≤ 8ms — strong evidence already: min 7.8ms under load 12–16, better than the pre-branch best; the machine had no quiet window during final verification)

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
pretable	Ignored		Aug 19, 2026 5:43am

[github-actions]

Vercel preview ready

Preview: https://pretable-a8ib39q95-cacheplane.vercel.app
Commit: 396ec3b78b51f92738bb82f1b531232fcc52fea4

_{Updated automatically by the deploy-preview job.}