Input latency profiler + recorder + baseline (review-only)

[hitpopdimestop]

Review-only PR against my fork so the profiling work is easy to read as a single diff. Not intended for upstream as-is (upstream tracking is issue superdoc#3842).

What this branch adds

An opt-in, zero-by-default input-latency profiler for the presentation editor, plus a reproducible browser recorder/comparator and a captured baseline. No product behavior changes unless profiling is explicitly enabled.

Instrumentation (packages/super-editor)

• InputLatencyProfiler (presentation-editor/performance/) — created only when SD_DEBUG_INPUT_LATENCY / __SD_DEBUG_INPUT_LATENCY__ is set. Correlates beforeinput → transaction → coalesced render → per-phase layout → DOM paint → next frame and emits standard User Timing measures. When disabled it allocates nothing (one null check at each call site).
• PresentationInputBridge — one synchronous hook for an eligible visible beforeinput (including input that lands on the focused hidden editor).
• PresentationEditor — records the transaction, render coalescing, per-phase timings, DOM paint, and the following animation frame; cleans up on destroy.
• Next-frame time is read with performance.now() in the rAF callback (not the frame-start argument), so there are no physically-impossible next-frame < DOM-paint samples.

Dev tooling (devtools/input-latency/, dev-only, alongside word-benchmark-sidecar)

• profile.mjs — Playwright recorder: fixed protocol (1 warm-up + 5 measured runs of 40 chars @100Ms across normal / 4×-CPU / layout-off), fixture SHA verification, full trace/discard metadata, strict completeness validation before write.
• report.mjs / report.test.mjs — deterministic statistics, fingerprinting, comparison, and report-completeness validation (unit-tested).
• compare.mjs — strict comparison CLI with a materiality rule (max(5ms, 2×MAD)).
• README.md — exact record/compare workflow. Raw reports live under results/ and are gitignored.

How to review

• The packages/super-editor changes are the instrumentation surface; check the opt-in gating and the disabled-path no-ops.
• devtools/input-latency/ is standalone tooling; report.test.mjs covers the pure logic.
• Findings and recommended next steps are written up in perf: first edit re-measures the whole document — measure cache is keyed by a document-wide font signature superdoc/docx-editor#3841 (measure-cache) and perf: input-to-paint latency on large documents — profiling harness, baseline, and where the time goes superdoc/docx-editor#3842 (latency baseline).

Notes

• pnpm-workspace.yaml bumps playwright catalog to ^1.60 (Node 24.16 extract-zip install hang); pnpm-lock.yaml picked up some incidental @types/node drift that would be cleaned before any upstream submission.
• Package type-check has a large pre-existing failure baseline in this repo; these changes add no new type errors.