fix: use monotonic time for startup observation

[ooiuuii]

Why

Startup observation measures elapsed wait with Date.now. A forward clock adjustment can return ok:true before a child exits nonzero; a backward adjustment extends the awaited startup HTTP response. Internal workspace creation consumes this result to show or clear its startup error.

Approach

Use performance.now for the four elapsed-clock reads in the existing autostart helper. Keep observation durations, silent-start extension, exit/cancellation semantics, wall-clock timestamps and receipts unchanged. One owner file, no new abstraction or dependency.

Validation

New real HTTP/SQLite/native-PTY diagnostic on 2026-10-09, Linux Node24.19.0, main f8f1343:

• Corrected RED 02:40:17–02:40:24 UTC: 3 failed/6 passed. Forward +60s returns success in34ms before child exits23; continuing child also returns early. Backward -2s makes normal observation2808ms instead of808ms
• GREEN 02:40:37–02:40:43: all9 scenarios pass. Early exits detected with steady/forward/backward clocks; continuing children observe810/812/810ms; real HTTP stop settles promptly
• Actual createApp routes, store, child and HTTP requests; only in-process Date.now is offset after an actual waitForRunExit. No system-clock change, fake PTY/fetch or provider calls
• 24 existing focused autostart/cancellation/error tests passed; pnpm install --frozen-lockfile, pnpm check, pnpm build, git diff --check and clean-main patch application passed
• Current CI-selected affected closure completed unchanged: 593 passed,1 failed,1 skipped across105files. The sole lifecycle-hardening descendant-cleanup assertion was freshly reproduced on clean main (10pass/1samefail). This gate is NOT green
• One scoped review found no blockers; no subsequent source edits

The runnable nine-scenario server self-check is retained outside the production diff, under AGENTS' current runnable-demo/new-test guidance; this is not a UI-test exemption. Earlier diagnostic mistakes about stop status/timer cleanup were corrected and retained; only the corrected RED is counted.

Not run: full default pnpm test (current classifier selects affected closure), whole-project tsc, package/release/native matrix, UI walkthrough, providers, Windows/macOS/Node22/26, full clean-main suite. Silent4000ms extension and legacy polling fallback have no direct clock-adjustment case. Existing cooperative observation is not a hard real-time guarantee. Remote CI remains unverified; read-only-home warnings retained without permission changes.