perf(highlight): cut tab init from ~20s to ~5s on heavy pages

[softpudding]

Summary

Tab init's two heaviest phases on dense pages (finviz.com/screener) were the in-page DOM scan (~10s) and the background pagination (~4s). Both share the same shape — per-element loops that re-do work the previous step already paid for. This PR cuts them with caching and a spatial index, no algorithmic risk to correctness.

Phase	Baseline	After	Δ
In-page scan (finviz, ~349 candidates)	6537 ms (10 s before any caching)	585 ms	−91%
Pagination	4106 ms	300 ms	−93%
End-to-end tab init	17.8 s	5.0 s	−72%

On the 16 deterministic mock sites, aggregate latency drops 12–47%, biggest wins on the heavier pages (bluebook −47%, techforum −31%).

Commits

• `96efe39` perf(highlight): cache layout reads + spatial-index pagination
• `3299e5f` perf(highlight): cache resolve-phase classifiers + emit perf breakdown
• `01af8f6` set extension auto reload timeout to 40s

What changed

`extension/src/commands/highlight-detection.injected.js`

• `withScanLayoutCache(fn)` monkey-patches `Element.prototype.getBoundingClientRect`, `SVGGraphicsElement.prototype.getBoundingClientRect`, `Document.prototype.elementsFromPoint`, and `window.getComputedStyle` for the duration of one scan. Backed by per-scan WeakMap/Map caches; restored in `finally`.
• Per-scan WeakMap memoization for resolve-phase classifiers that are pure functions of element + DOM state: `getSemanticClickableSignal`, `isClickableCandidate`, `getBaseClickableSignal`, `hasExplicitClickableAncestor`, `getElementTextForDetection`, `getElementSearchText`. The resolve phase was 6.1s of the 6.5s scan; this drops it to ~50ms.
• Tag pre-filter for inert elements (script/style/meta/...) before any layout read.
• `_perf` field on the response payload with `scan_ms`, `scan_stats`, `scan_times`, `pagination_ms`, `screenshot_ms`, `consistency_ms` — for tooling/profiling without parsing console output.

`extension/src/utils/collision-detection.ts`

• `SelectedSpatialIndex` (96 px grid) keyed on `union(bbox, labelBBox)` of placed elements.
• `isPlacementFeasible` now iterates only nearby placed elements via `nearbySelectedFor`, which queries by `inflate(union(candidate.bbox, candidate.labelBBox), CLEARANCE)` — covering all four collision tests (label-label, bbox-bbox, candidate-label-vs-selected-bbox, candidate-bbox-vs-selected-label).
• `chooseLeastBlockingPlacement` uses an "influence rect" to skip re-evaluating spatially-far future candidates whose feasibility a hypothetical placement cannot affect.

`extension/src/background/index.ts`: thread `_perf` into the response.

`extension/vite.config.ts`: dev-reload wait bumped 10s → 40s to cover MV3 service-worker keepalive cycles.

Why caching is safe

The scan runs in a single synchronous `Runtime.evaluate` call. No `await`, no `setTimeout`, no MutationObserver callbacks. Page JS cannot run concurrently. Layout/style cannot mutate mid-scan, so cached rect/style/EFP values are guaranteed fresh. Originals are restored in `finally` regardless of throw.

Correctness verification

• 181 / 181 extension unit tests pass (incl. `highlight-placement`, `highlight-integration`, `hit-test-visibility`).
• Strict integration check (selector + type + labelPosition + bbox + element ORDER) passes on all 16 deterministic mock sites — same elements, same labels, same visual order.
• Real finviz returns identical 336 elements / 6 pages / 138 on page 1 across baseline and optimized runs.
• Codex code review consulted on both the plan and the diff; concerns surfaced (NaN guard in `forEachCell`, restoring the original `elementsFromPoint` to the prototype rather than to a bound wrapper) were addressed before commit.

Test plan

• `bun test` in `extension/` — 181 tests pass
• Open the dev extension, init a tab to https://finviz.com/screener.ashx, confirm `_perf` shows `scan_ms < 1500`, `pagination_ms < 500`, total init < 7s
• Repeat on at least 3 mock sites under `eval/` (finviz_mock, bluebook, techforum) — confirm element counts and visual order match baseline

Stack

This is PR 3 of 3:

1. `feat/image-input-and-file-upload` → `main` (feat: image input for prompts + upload_file browser action #61)
2. `feat/ob-routines-skill` → `feat/image-input-and-file-upload` (skill(ob-routines): record → compile → replay browser routines #62)
3. This PR → `feat/ob-routines-skill`

After #61 and #62 merge to main, retarget this PR to `main`.

🤖 Generated with Claude Code