fix(ios-profiler): make xctrace export async so native-profiler-stop can't trip the tool-server health check

[filip131311]

Problem

On the newest main (0.12.1), native-profiler-stop on an iOS 26.5 sim returns:

Missing or invalid Authorization header. Tool-server requires `Authorization: Bearer <token>`…

…even though the trace export actually completes and the XML lands on disk. Reproduced profiling Settings on a 26.5 sim.

Root cause — event-loop starvation, not a profiler bug

1. native-profiler-stop runs four xctrace export passes (TOC discovery, CPU, hangs, leaks) through execSyncWithTimeout → child_process.execSync, which blocks the Node event loop for the entire export.
2. iOS 26.5 forces the host-wide --all-processes capture fallback (the xctrace --device deadlock workaround, feat(ios-profiler): pluggable capture strategy with --all-processes fallback for the xctrace --device deadlock #380). That export is ~44 MB and the whole stop takes ~34 s (durationMs: 34083 in mcp-calls.log).
3. While the loop is frozen, the tool-server's GET /tools health endpoint can't answer. The MCP client's ensureToolsServer() gates every call on isToolsServerHealthy() with a hard 2 s timeout:

   const healthy = await isToolsServerHealthy(state.port, host, 2e3, state.token); // 2s
   if (healthy) return {…reuse…};
   await clearState();
   const token = generateToken();   // ← respawn + NEW token

4. The busy-but-alive server fails the 2 s check → the client declares it dead, spawns a second tool-server and rotates the auth token in ~/.argent/tool-server.json → the in-flight stop is 401'd. Confirmed live: one MCP client, two tool-server.cjs processes.

The recording start path already uses async spawn + await, which is why it never trips this; only the export path was synchronous.

Fix

Convert the four export passes to async (exec + await) so the event loop stays free to answer health checks while xctrace runs. Carried over verbatim onto the async wrapper:

• timeout: 60_000 — still caps a genuinely-stuck xctrace.
• maxBuffer: 256 MiB — the ENOBUFS fix (fix(ios-profiler): set maxBuffer on xctrace export to stop ENOBUFS killing every export #382); exec honors it.

promisify(exec) is resolved per-call (not at module load) so test suites that vi.doMock("child_process", …) without an exec export still import the module — matching the original lazy execSync usage.

This does not change stop latency (the export is genuinely slow) — it makes it non-blocking, which is the actual bug. Concurrent calls (health checks, screenshots) now work during a long stop.

Not addressed here (follow-ups)

• MCP respawn hardening: ensureToolsServer() still rotates the token / abandons in-flight requests against a still-alive PID the moment a health check misses. Async export removes this trigger; the brittle respawn behaviour is a separate fix.
• Analyze path: if native-profiler-analyze parses the 44 MB XML synchronously it can block the loop similarly — worth checking next.

Testing

• tsc --noEmit clean
• eslint clean on changed files
• Full tool-server suite: 1314 tests pass (incl. test/ios-instruments/, 34 tests). Updated the exportIosTraceData mock from mockReturnValue → mockResolvedValue.