[Avicennasis]

FreeToken currently inherits uvicorn's 5-second HTTP idle timeout. Add --http-keep-alive-s, default 360 seconds, and pass it to uvicorn in both normal and terminal-shell startup. Document that it must exceed the longest client pool idle timeout; a header-blind client pool can otherwise race the server's close and lose an unretried POST.

The idea comes from Infernix PR #55, whose revision was cherry-picked as 9ee5e543, and the maintainer's 360-second follow-up. The subsequent buffered-close change corrects an earlier claim: Bun probes pooled sockets, but a FIN can arrive after the probe. This configuration change copies no donor code; both projects use Apache-2.0. It does not add Infernix's Keep-Alive advertisement, close buffer or independent HTTP-client ceiling. Uvicorn's limit_concurrency remains unset because it couples connections/tasks to request admission.

Verified on Linux x86_64, Python 3.12.3, CPU-only torch 2.14.1+cpu, uvicorn 0.54.0. From the repository root:

export PYTHONPATH=python
python -m pytest tests/server/test_api_server.py -q
python -m pytest tests/server/ -q
python -m pytest tests/ -q --continue-on-collection-errors

The regression was written first: 4 failed on the base (both startup paths used 5 seconds; the override flag was unknown), then 4 passed after the change. It inspects real uvicorn.Config objects without binding a socket. The server subsystem has 660 passing tests. Full suite on the branch: 1669 passed, 20 failed, 919 skipped, one warning. On the base: 1665 passed, 20 failed, 919 skipped, one warning. All 20 failing IDs are identical and are the known CPU-only environment failures. python -m freetoken.cli serve --help and git diff --check also passed.

Bun 1.4.4 ran two POSTs with a 10-second idle gap through FreeToken's actual startup with a CPU-only echo route and no model. Baseline returned HTTP 200 twice on different connections; both the new default and explicit --http-keep-alive-s 360 returned HTTP 200 twice on the same connection. This verifies the idle lifetime, not the timing race: the lost POST on main was NOT reproduced. Real small-model ft serve before/after, GPU/checkpoint loading, 300-second pool expiry and live terminal-shell interaction are NOT VERIFIED.

Based directly on 628b88f and does not depend on #334, #459, #565, #593 or draft #656; their file lists do not overlap. #592 overlaps only in docs/cli.md, in the separate multimodal section. Its media flags are already on main, so this argument is added beside --port. No merge-tree stacking check was run.

Maintainer question: is 360 seconds an acceptable HTTP idle default for Bun/opencode deployments? Operators serving longer-lived header-blind pools must choose a larger value with margin.