fix: wait for daemon health before reclaiming occupied port

[s09x]

Summary

• Add a bounded health grace window before reclaiming occupied daemon ports.
• Reuse an already-starting daemon only after /health returns Hindsight's initialized health payload.
• Preserve stale listener reclaim behavior after the grace window expires.

Reproduction

A concurrent daemon start can bind the profile port before /health is ready. The previous _clear_port behavior checked health once, saw the warming daemon as unhealthy/non-Hindsight, and moved into PID lookup/reclaim instead of waiting for the daemon to finish starting.

The regression coverage reproduces that race deterministically without real services or private data:

1. _is_port_in_use(port) returns True.
2. /health responses are simulated as 503, 503, then 200 with {"status": "healthy", "database": "connected"}.
3. Previous behavior returns False after the first health probe and calls _find_pid_on_port.
4. Patched behavior keeps polling within the grace window, returns True when the daemon becomes healthy, and does not call _find_pid_on_port or _kill_process.

Fix

• Add configurable, bounded port-health grace settings:
  • HINDSIGHT_EMBED_PORT_HEALTH_GRACE_TIMEOUT
  • HINDSIGHT_EMBED_PORT_HEALTH_CHECK_INTERVAL
• Add health helpers that require the initialized Hindsight health payload, not just any HTTP 200 response.
• Use monotonic time for the grace deadline.
• Treat listeners that disappear during the grace wait as already cleared.
• Reclaim stale/unhealthy listeners only after the grace window expires.

Test Plan

• uv run pytest tests/test_daemon_client.py::TestClearPort -q
• uv run pytest tests/test_embed_manager.py -q
• uv run pytest -q
• uv run ruff check hindsight_embed/daemon_embed_manager.py tests/test_daemon_client.py
• git diff --check --cached -- hindsight-embed/hindsight_embed/daemon_embed_manager.py hindsight-embed/tests/test_daemon_client.py

[s09x]

The failing check appears to be a transient dependency download issue in verify-generated-files, not a code/test failure.

The job failed while Cargo was updating the crates.io registry:

failed to get `h2` as a dependency
unable to update registry `crates-io`
download of 2/h2 failed
curl failed
[16] Error in the HTTP2 framing layer

The local checks I ran before opening the PR completed successfully:

• uv run pytest tests/test_daemon_client.py::TestClearPort -q
• uv run pytest tests/test_embed_manager.py -q
• uv run pytest -q
• uv run ruff check hindsight_embed/daemon_embed_manager.py tests/test_daemon_client.py
• git diff --check --cached -- hindsight-embed/hindsight_embed/daemon_embed_manager.py hindsight-embed/tests/test_daemon_client.py