fix(pool): healthcheck timer must not block graceful worker shutdown (true-async/server#93)

[EdmondDantes]

Problem

Under a debug build, hot-reloading workers can trip a scheduler assert:

Assertion failed: (zend_async_reactor_loop_alive_fn() == false && "The event loop must be stopped"),
function fiber_entry, file scheduler.c, line 2041.

Reproduced deterministically: a worker bootloader that warms an async PDO connection pool with PDO::ATTR_POOL_HEALTHCHECK_INTERVAL > 0, then reload(). With the interval set to 0 (no timer) the reload is clean — isolating the pool's healthcheck timer as the cause.

Root cause

pool_start_healthcheck_timer() created the background heartbeat timer as a normal reactor event, so libuv_timer_start incremented active_event_count. On graceful shutdown start_graceful_shutdown() cancels coroutines, but nothing owns/closes this timer — the pool is only disposed later at PDO object teardown. So ZEND_ASYNC_REACTOR_LOOP_ALIVE() (active_event_count > 0 && uv_loop_alive) stayed true while the scheduler fiber finalized:

• debug: trips the loop-alive ZEND_ASSERT in fiber_entry.
• release: an idle pool needlessly keeps the retiring worker's loop alive.

The ZEND_ASYNC_EVENT_F_HIDDEN flag exists for exactly this — its doc names "Background timers (e.g., garbage collection, health checks)" as the canonical case — but the pool timer never set it.

Fix

Mark the healthcheck timer HIDDEN right after creation. The timer still fires (HIDDEN only gates the active_event_count accounting, not uv_timer_start); it just no longer counts as application work for loop-alive / deadlock detection.

Verification

• Repro (pool + interval=30 + reload ×N): was exit 134 with assert → now 0 asserts, clean shutdown, all reloads succeed.
• server hot-reload phpt suite (051–054): 4/4 PASS.
• PDO pool phpt suites (pdo_pool_*, pool_*): 0 failed (17 passed, 8 skipped — need mysql/pgsql).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!