test: close the five highest-value gaps from the test audit

[tt-a1i]

Problem

Five tests were protecting the right failures, but the way they checked those failures was brittle or never reached the failure.

• The web performance fixture (test(web): 修复性能 fixture 的 Session 路径身份，恢复性能测量入口 #840) replaced the session id and left the host session path in place. The UI stayed on reconnect, so the 11 cases timed out before measuring anything. The suite is not part of required CI.
• Windows setup-config lock recovery (Flaky CI tests: session-history E2E anchor, Windows git-info refresh, Windows setup-config lock timeout #829) asked PowerShell for the process start time inside a 1s query budget. A cold PowerShell start made the query unknown, so a dead owner could not be recovered before the 5s fail-closed deadline.
• The Antigravity lifetime test used one 5ms timeoutMs for both the first SSE event and the rest of the stream. On a slow run the metadata event was still the first-event timeout, so the test expected "next SSE event" and observed "first SSE event" (Flaky CI tests: session-history E2E anchor, Windows git-info refresh, Windows setup-config lock timeout #829).
• tests/web/session-worktree.test.ts compared raw git index bytes. With core.fsmonitor or core.untrackedCache, git appends index extensions and the bytes change even when the staged content does not.
• scripts/provenance.mjs treated a path containing openpi as an OpenPI source. A checkout whose directory is not named that, including this /workspace tree, was reported as no match even when pi list showed only that checkout.

Value

Each check still fails the original bad result, and the false failures stop blocking the measurement or the recovery they exist to protect. The performance suite can run. Provenance accepts the real checkout and still rejects zero, foreign, or multiple OpenPI roots. Lock recovery reads a Windows start time from WMI only. A missing time stays unknown and the 5s lock expires without taking the lock.

Approach

• The performance snapshot sets currentSessionId and currentSessionPath to the same synthetic session, and assertPerformanceSessionIdentity throws before the route is fulfilled when those disagree with the session list or the selected session. An identity error is not swallowed by the route abort path. The suite is not added to CI.
• Windows start time comes from one cscript.exe //Nologo query of Win32_Process.CreationDate, read as the CIM datetime string and truncated to unix milliseconds. Redirected cscript output is decoded as UTF-16 LE when it has a BOM or NUL-padded ASCII; a UTF-8 misread stays unparseable. The script path is resolved when the query runs, so Vitest can import setup config without a file URL. The JScript has no trailing commas, writes a failure token and quits outside try, and is not launched with //B. The query budget stays 1s. There is no PowerShell fallback and no second query command. A non-zero exit, timeout, or unparseable value is unknown ownership, and the lock deadline stays 5s. Linux still uses ps. The existing execFile failure injection still covers the fail-closed path, and on Windows that injection must be cscript.exe only.
• The product still applies timeoutMs to both Antigravity deadlines. The lifetime test holds Date.now until the metadata event has been parsed, then moves the clock on the third body pull (the response prefetches one chunk before that parse). A sibling test advances the clock before any payload is parsed and still expects "first SSE event".
• Session-worktree index checks use git ls-files --stage -z. The fork test still compares git status --porcelain=v1 -z.
• Provenance dedupes reported paths with realpath, counts a path as OpenPI when it is this checkout or its package.json name is @tt-a1i/openpi, and matches only when that set is exactly the checkout. Importing the module no longer runs pi list. The foreign-package test compares realpathSync.native of that package, so a Windows 8.3 temp path is not treated as a different source.

Validation

Linux x64, Node v22.23.3, Bun 1.3.14. Chrome 154 for the performance suite.

• bun run lint and tsc --noEmit passed after removing the PowerShell fallback. tests/extensions/shared/setup-config.test.ts passed on Linux: reused-PID recovery, the mocked query-failure lock wait (about 5.5s, no mutation, no PowerShell command), and the live WMI test skipped off Windows.
• bun run test passed on d00476b: Node 2639 passed / 9 skipped / 0 failed (2648 tests), Vitest 1520 passed / 98 files.
• Performance suite, not in CI: 11 passed in 16.8s. Bundle SHA-256 unchanged: 99577ba28cb347d38db1008979c9b30e08befa5baa34e0dfb09530afd2b4ecfe.
• Head ea256c6 removed the PowerShell fallback but the WMI script was invalid JScript (a trailing comma, hidden by //B). Windows CI on that head failed the live WMI read in 54ms and then the reused-PID recovery at the 5s deadline. Head fe57ee4 fixes the script. CI run 38865595760 on that head succeeded. Windows parallel 2/2 recorded Windows reads this process start time from WMI alone and a reused PID does not impersonate the dead lock owner as passing, and the query-failure case still waited out the deadline (ok 458 about 6.5s after ok 457).

No speedup versus PowerShell is claimed. Heartbeat liveness (#862) is out of scope.

Impact

• User-visible behavior: none for a normal session. A Windows config-lock recovery reads the process creation time through WMI only. A failed read does not take the lock.
• Model-visible context/tools: none.
• Runtime/lifecycle: setup-config lock ownership and the provenance gate. The performance suite remains opt-in.
• Persisted config/data: lock owner files still store processStartedAt. No config schema change.
• Compatibility or risk: the Windows start-time command is cscript only. The 5s fail-closed deadline and the 1s query budget did not change.

Fixes #840. Related to #829.