fix(dashboard): cache gateway PID status probes

[0disoft]

Summary

• Add a short-lived cached wrapper around get_running_pid() for read-only dashboard/status polling.
• Invalidate the cache when PID, runtime lock, or runtime status files change, and when the current process writes/removes PID or lock ownership.
• Use the cached probe from /api/status so high-frequency dashboard polling does not repeatedly open and lock gateway.lock.

Root cause

/api/status called get_running_pid() on every request. That path checks is_gateway_runtime_lock_active(), which opens gateway.lock and probes the advisory lock each time. Under active dashboard polling this churns file descriptors and can hit OSError: [Errno 24] Too many open files in the dashboard process.

Validation

• bash scripts/run_tests.sh tests/gateway/test_status.py tests/hermes_cli/test_web_server.py could not start on this Windows host due to the local Bash/CallMsi REGDB_E_CLASSNOTREG launcher error.
• .\.venv\Scripts\python.exe -m pytest tests/gateway/test_status.py::TestGatewayPidState -q -> 15 passed
• .\.venv\Scripts\python.exe -m pytest tests/hermes_cli/test_web_server.py::TestWebServerEndpoints tests/hermes_cli/test_web_server.py::TestStatusRemoteGateway tests/hermes_cli/test_web_server.py::TestGatewayBusyReadout -q -> 136 passed, 4 warnings
• ruff check gateway/status.py hermes_cli/web_server.py tests/gateway/test_status.py tests/hermes_cli/test_web_server.py -> passed
• git diff --check -> passed, with CRLF conversion warnings only

I also ran the full tests/gateway/test_status.py file directly on Windows. It now has 86 passed, 3 failed; the remaining failures are pre-existing Windows/local-environment failures unrelated to this PR (sleep executable missing for two process-start-time tests, plus an existing Windows path/cmdline assertion in test_runtime_status_running_pid_accepts_matching_profile_cmdline).

Fixes #53484

[tonydwb]

Code Review Summary

Verdict: Approved

Adds file-change-based caching for get_running_pid() dashboard polling. Uses threading.Lock for thread safety, file stat signatures (mtime_ns, size) for cache invalidation, and a 1-second TTL. Properly clears cache on PID file writes, lock acquisition/release, and stale cleanup. Includes thorough tests.

---

Reviewed by Hermes Agent

[kshitijk4poor]

Merged via #60884 (rebase) — your commit landed on main as 7d0ddbb with your authorship preserved. The TTL + file-signature cache design held up well under review (the signature check even covers a legacy direct-unlink path in --replace). Thanks for the EMFILE fix!