test(agent): steady-state 'ready' signal to kill the startup-timing flake class

[LiorFink00]

What

Closes #237. A whole class of macOS flakes this cycle shared one root cause: a test acted immediately after a weak readiness gate (a bare "armed?" check), racing the agent's sub-second startup — e.g. simulating an atime read before atime_poll had captured its baseline, so the read was captured as the baseline and missed.

Agent

Each watcher now emits a single unambiguous steady-state signal once it's fully up — workers spawned AND (for atime) baselines captured:

• mark_ready() — writes a ready marker + logs watcher ready, from watch_fifo / watch_mixed / atime_poll / watch_inotify. This also makes the ephemeral ready marker accurate in every mode (it was previously written in the main flow right after start_watcher returned — before the backgrounded watcher had armed, which is exactly the bug).
• mark_armed() — bumps an atime_armed generation after each baseline (re)capture, so a test can wait past a specific re-arm instead of a blind sleep (the arm→baseline window is a genuine race, agent: atime_poll can swallow a read in the arm→baseline-capture window #235).

Tests

A shared helper (tests/agent_ready.py: wait_ready / wait_atime_gen) replaces the ad-hoc "armed?" gates + sleep(2) settles in the four flaky tests: atime rearm, mixed plant+fire, ephemeral terminate, live-sync reconcile.

Verification

• test_atime_sensor_is_rearmable: 0/15 flaky before under local stress → 15/15 after.
• All four migrated tests: 10/10 stress rounds (40 runs) clean.
• Full suite 375 passed, 1 skipped; ruff + shellcheck clean.
• Real-endpoint behavior unchanged — a real reader hits a steady-state agent, never this startup window.

This should end the recurring macOS flake that has repeatedly blocked merges this cycle (it forced re-runs on #268, #272, #258, and others).

🤖 Generated with Claude Code