Evict terminal chat runs to bound ChatRunRegistry memory

[TYRMars]

Summary

Fixes #50. ChatRunRegistry never evicted terminal (completed/failed/cancelled) runs from its in-memory inner map, so memory grew unbounded over the lifetime of a long-lived server — one retained entry per distinct conversation_id ever run, each holding up to MAX_EVENTS_PER_RUN buffered records plus a live broadcast channel.

Fix

Lazy eviction on the existing mutation paths — no background task:

• TTL: terminal runs are dropped TERMINAL_RETENTION_MS (5 min) after they finish, leaving a window for late WS reconnects to replay the final frames and observe channel closure before the state disappears.
• LRU backstop: retained terminal runs are capped at MAX_RETAINED_TERMINAL (256), dropping the oldest by terminal timestamp — guards against a burst of conversations all completing inside the TTL window.

Active runs (terminated_at == None) are never counted against the cap or evicted. The terminal timestamp is stamped exactly once on the first terminal transition via mark_terminal, which also drives the one-shot abort-handle cleanup. The sweep (prune_terminal_locked) runs from try_start and from the terminal branch of push_frame / update, all already under the inner write lock, so it adds no new locking and only fires on run start / completion (not per token-delta).

Tests

Added coverage in chat_runs::tests (via a force_prune test seam that injects the clock so TTL is testable without sleeping):

• terminal_run_is_evicted_after_ttl — kept within the window, swept past it
• active_run_is_never_evicted — even with an absurd clock
• cancelled_run_is_evicted_after_ttl
• terminal_runs_are_capped_by_lru_backstop
• restarting_an_evicted_conversation_succeeds

cargo test -p harness-server chat_runs:: (11 passed) and cargo clippy -p harness-server --all-targets -- -D warnings both green.

---

Generated by Claude Code