fix(driver): wait for network and render quiet

[amazon7737]

A single pending fetch could reach network idle before its response, and deferred rendering could leave a skeleton in the next observation. Continuous polling could instead consume the entire timeout.

Chrome now requires a quiet interval across nonexcluded request identities/statuses and document mutations. Pending requests keep the wait active; a new response or render restarts the interval. The independent, expiring DOM observer preserves exact-reference guards. One deadline also bounds hung tool calls and the final sleep.

Consumers can configure ChromeDriverOptions.settle.ignoreRequests with URL substrings. Those defaults apply to engine and internal waits, with individual per-call overrides. Excluded requests remain cumulative evidence and still fail assertions; this does not change benign or frozen scenarios.

Closes #175.

Validation:

• Typecheck, build, full workspace/benchmark-script/demo tests, boundaries, language and diff checks pass.
• All 30 Chromium probe/reference tests pass.
• Real MCP 1.8.0 document-reference and cumulative-network regressions pass, including deterministic replay.
• npm run test:settle uses committed local fixtures and public Driver/engine APIs. It checks a pending fetch followed by deferred rendering, polling exclusions with an unrecovered 503 still failing the built-in guard, continuous DOM mutation reaching its cap, and a server-verified commit with zero model calls.
• One local baseline comparison against develop 31cfed0, with a 400ms quiet interval: the original wait returned in 423ms while the fetch was pending; the fix waited 2764ms through completion and rendering. Excluded polling took 543ms instead of 2212ms. These are fixture observations, not general latency estimates.
• The MCP CI job now runs the new fixture and preserves its JSON report.

Settling remains best effort. Missing DOM measurements fall back to network waiting; numeric response status does not prove a streamed body finished. CSS-only changes, shadow/frame rendering, continuous DOM updates and work scheduled beyond the quiet window still require explicit expect/waitFor checks. No hosted model was used.

[cairn-pingu]

🐧 Performance briefing

• Package tarball: +1128 B (+0.48%). Browser gzip: 0 B (0%).
• Observed replay medians: navigation: +24.4%; form: +30.64%; stateful: +23.82%.
• Execution across both revisions: 24/24 passed; engine LLM calls: 0; observed LLM calls: 0.
• Timing includes server/browser startup and awaited cleanup; small samples and CI noise do not prove an improvement or regression. Residual order bias may remain.

Detailed measurements

PR performance comparison

Before: 31cfed0
After: 797391e

All measured attempts passed with zero engine-reported and observed LLM calls.

Package size

Size	Before bytes	After bytes	Delta bytes	Change
Package tarball	237067	238195	+1128	+0.48%
Package unpacked	875969	880980	+5011	+0.57%
Browser bundle	69557	69557	0	0%
Browser bundle gzip	24067	24067	0	0%

Replay elapsed time

Tier	N before / after	Median before ms	Median after ms	Delta ms	Change	Status
navigation	4 / 4	2595.33	3228.57	+633.23	+24.4%	higher median (informational)
form	4 / 4	3789.6	4950.6	+1161	+30.64%	higher median (informational)
stateful	4 / 4	4874.49	6035.48	+1160.99	+23.82%	higher median (informational)

Tier	Failures before / after	Engine LLM calls before / after	Observed LLM calls before / after
navigation	0 / 0	0 / 0	0 / 0
form	0 / 0	0 / 0	0 / 0
stateful	0 / 0	0 / 0	0 / 0

Raw measurements and job log

[solp721]

LGTM