[coaxk]

Closes #626.

What the measurement showed

Step 1 of #626, measured with a synthetic media_probe (2 audio + 3 subtitle streams per row), Python 3.12, --cpus 2 (the maintainer instance's limit), loop_lag at its default 0.5 s threshold:

Coverage refresh (build_coverage, 3 builds per size)

rows	all_entries() in the build	longest loop gap	loop_lag warnings
6,375	106-183 ms	16-71 ms	0
40,000	991-1,240 ms	142-298 ms	0

The hydration already runs in asyncio.to_thread, and CPython hands the GIL back every 5 ms, so the loop is slowed, not frozen. The issue's acceptance (no loop_lag warning from a coverage refresh at 40k) already holds on main, so no caching, lazy hydration or chunked decode is added there: each would add memory bounds or invalidation logic to remove a gap that is under the threshold.

GET /api/probe/library (the Library Probe tab) called all_entries() directly on the event loop:

rows	loop blocked per call	loop_lag warnings
6,375	107-183 ms	0
40,000	1,038-1,484 ms	3 of 3 calls before, 0 of 3 after

The warning's stack named routers/probe.py library -> probe_store.py all_entries.

Change

• The route's body (hydration plus per-row classification and filtering) moves into _library_payload and runs in asyncio.to_thread. Response shape and every filter are unchanged.

Tests

• tests/test_probe_library_off_loop_626.py: all_entries() and the per-row classification must never run on the loop. Fails on main; a mutant that moves only the hydration off-loop fails the classification assert.
• Same file pins limit vs total_cached, only_without_eng (partial English counts as English), combined filter_text + sub_kind, and the empty store. These had no test before; they pass on main and after.
• tests/test_library_probe.py unchanged and green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ShzWm3urD55AUk5yrSgXsi