fix: reap dead-child sessions even when --session-idle-ttl is unset

[shigechika]

Summary

Closes #385.

SessionRegistry.reap_idle() already reaps two independent things — a dead child (backend.closed, unconditionally, TTL or not) and one idle past the TTL (only when idle_ttl > 0) — but start_reaper() gated starting the thread at all on the TTL alone. In the default deployment (--session-idle-ttl unset, 0) the thread never started, so the unconditional dead-child sweep had no production caller: a long-lived gateway on default settings accumulated dead-child sessions with no way to shed them, eating into --max-sessions until new initialize calls got 503 against slots that were all corpses.

Fix

start_reaper() now always starts (no-op only on a second call): the tick interval favors the idle TTL when one is configured, and falls back to sweeping for dead children alone, at least once a minute (_MAX_REAP_INTERVAL_SECS), when it is not.

This mirrors the identical fix already applied to the modern pool's analogous reaper (#390 R3F1, ModernBackendPool.start_reaper) — separate "why does the thread run" from "what does each tick evict", since reap_idle was always correct on the latter.

Deliberately out of scope

The modern pool's own reaper still has a residual version of this gap — its start condition needs --modern-idle-ttl or the MRTR bridge enabled, so a pure-default deployment (neither set) still never starts it. That's the same class of bug, just not this issue's literal scope (#385 was explicitly split out of #383 to stay a narrow, provable fix). Happy to file a follow-up if wanted.

Test plan

• pytest tests/ -v — 1993 passed, 3 skipped
• pytest tests_integration/ -v — 42 passed
• ruff check . / ruff format --check . — clean
• New tests: the reaper thread now starts with idle_ttl=0; it sweeps a dead-child session even with no TTL configured (via a monkeypatched fast tick interval, so the test doesn't wait out the real 60s fallback); a merely-idle-but-alive session is left alone in that same configuration, proving this didn't silently enable full idle eviction

https://claude.ai/code/session_019Yjo4vhXdQAbA6xHPW9NdS

[github-actions]

AI review (gpt-5.6-sol)

Verdict: no blocking findings.

Scope: new commits since f855e2c (full PR diff supplied to the model as context).

No findings clear the reporting bar this round.

Findings ledger (all rounds)

• R1F1 [fixed/advisory] — Docstring falsely calls modern and legacy reaper fixes identical
• R1F2 [fixed/advisory] — Negative reaper test passes if thread dies
• R2F1 [fixed/advisory] — Initial count assertion races the running reaper

_{Advisory per-push review (round 3) generated by ai-review — verify findings before acting.}