Daemon can keep a dead writer lock, then run for hours with ready:false while every query times out #2918

What happened

On Windows 11 with the MCP lane serving CodeGraph 1.6.2, codegraph_explore began timing out on one
project while the CLI still reported a version and the project was indexed (12,451 files).

The .codegraph directory contained:

writer.pid  ->  {"pid":58188,"mode":"fallback","startedAt":...,"ready":false}   (pid 58188: not running)

Clearing that and reconnecting left a second state:

writer.pid  ->  {"pid":20632,"mode":"fallback","startedAt":...,"ready":false}   (pid 20632: running, ~28 h)

The process was alive, had been up roughly 28 hours, and never reported ready:true. Every query against
that project timed out until the daemon was killed.

Why it is hard to recover from

• The dead-pid case is recoverable (codegraph unlock, or deleting the stale artifacts).
• The live but never-ready case is not: a pid-liveness check says "a writer owns this lock", so nothing
  clears it, and the daemon keeps holding the project while answering nothing.
• A client retry spawns another daemon against the same lock, which makes it worse.

Reproduction sketch

1. Start the daemon on a project, then kill it with SIGKILL-equivalent (taskkill /F) so writer.pid keeps
   a dead pid.
2. Start a new daemon on the same project; observe it either wedge on the dead lock or come up with
   ready:false.
3. Query over MCP: the request times out rather than returning an error that names the lock.

We did not reproduce this from a clean upstream install in isolation; it was observed on a lane build of
1.6.2 that also carries unrelated local patches, so the exact trigger may differ. The lock handling and the
ready flag are upstream behaviour.

What would fix it

Either of:

• Write the pid only after the writer is ready, or add an ownership token with a heartbeat/TTL so a
  stale record is distinguishable from a live writer.
• Treat ready:false past a threshold as stale: clear the lock (and stop the owning daemon) when the
  writer has not become ready for N minutes, and say so in the log.
• At minimum, make the client failure name the lock file and the recorded pid instead of a bare timeout.

Environment

• CodeGraph 1.6.2, Windows 11, MCP lane (npm-shim.js serve --mcp)
• Project: a ~12k-file TypeScript monorepo
• Recovery that worked: codegraph unlock, kill only the daemon process whose command line contains the
  lane path, then one query

Tested upstream

CodeGraph 1.6.2, Windows 11, MCP lane, 2026-10-27; observed on a 12,451-file TS monorepo