Requests over ~4 KB hang for 30 s and return 500 (stale lf_toolkit pin)

[peterbjohnson]

The pinned lf_toolkit commit truncates IPC messages at 4096 bytes, so the Python worker never receives a complete request and never replies. Shimmy gives up after 30 s.

Impact: any evaluation with more than roughly 45 notes fails. The platform sends response/answer as JSON strings, so most real pieces exceed the limit. This affects the deployed staging function today.

Reproduce

docker run -d --name compare-music -p 9000:8080 <image>
# body with ~100 notes
curl -X POST http://localhost:9000/ -H "content-type: application/json" -H "command: eval" -d @big_body.json
# -> HTTP 500 {"error":{"message":"error sending rpc request: context deadline exceeded"}} after 30 s

Cause: NewlineStreamIO.read() in lf_toolkit/io/stream_io.py stopped at the size hint (4096) instead of reading to the newline delimiter.

Fix: upstream fixed this in lambda-feedback/toolkit-python#10 (19 Aug 2026). poetry.lock pins commit 8a687d3 from 29 Jul 2026, before the fix. Run poetry update lf_toolkit and commit the lock.

Verified locally: after the bump, a 289-note request returns in 0.14 s.