feat(bench): measure main-thread blocking on the interaction path

[blove]

Closes #458. Step 1 of the #452/#457 sequence — instrument before tuning, so the next change (the slice-budget lever) can prove it didn't regress main-thread blocking rather than assert it.

What

measureRowSetChange now runs a longtask PerformanceObserver and emits post_interaction_long_tasks_count / _ms. Until now the interaction path collected no blocking metric at all — the scroll and updates paths already did — so a synchronous engine blocking for its whole sort and a cooperative engine that never blocks were indistinguishable in every published interaction table.

Three deliberate choices:

• Attached at the trigger, not at function entry. A long task during the pre-trigger quiet wait is mount tail; charging it to the interaction poisons the metric. The trigger is the interaction — a blocking sort runs inside it.
• Push-based, so unlike the height-error walk removed in fix(bench): measure row-height error on recorded frames only #456 this adds no per-frame DOM work to the window being measured.
• Absent, not zero, when the host can't observe long tasks — same honesty rule as the row-height error.

First measurement

S2 sort, 3 repeats each, quiet machine:

	latency	long tasks
pretable @ 3k	~58ms	0
tanstack @ 3k	~33ms	0 (its ~15ms block is under the 50ms threshold)
tanstack @ 50k	~33ms	1 × 62ms, 3/3 runs, sd <1ms

So the honest comparative statement is finally expressible: at target scale TanStack reports a 33ms latency while blocking the main thread for 62ms per sort. (pretable at 50k is #457 — it never blocks but currently can't complete inside the window.)

Verification

The test stubs PerformanceObserver, plays synthetic entries at three moments — during the run's own quiet wait (via the frame stub's onFrame, not merely before the call — that distinction is what catches an observer attached at entry), inside the trigger, and after completion — and asserts count, total, and teardown.

Mutation-checked: attach-at-entry, never-disconnect, and metrics-dropped each redden the test. 156 bench tests, 16 bench-runner tests, typecheck, lint, format.

🤖 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 17, 2026 3:26am

[github-actions]

Vercel preview ready

Preview: https://pretable-87d86o85q-cacheplane.vercel.app
Commit: 284245b7fd22b43705ad0a92d595cdf2b51400b1

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