[Fix] Reject a non-finite timeout_s passed in code; real-server deadline evidence for #36

[chaimaerachdi]

Follow-up to #36.

1. Fix (from the [Feat] Make the decision deadline configurable for slower local System One servers #36 review): S1A_DECISION_TIMEOUT_S rejects nan, inf and values <= 0, but a deadline passed in code (JevModel.from_env(timeout_s=...)) skipped that check. JevDecisionsClient now applies the same guard and raises MODEL_SERVICE_CONFIG_ERROR naming timeout_s. New test: 0, -3, nan, inf, -inf. (c0b20d7)

2. Real-server evidence for [Feat] Make the decision deadline configurable for slower local System One servers #36, no injected delay: the Flights agent at the default 5 s, at S1A_DECISION_TIMEOUT_S=30 and at =1, on one NVIDIA A40.

• docs/results/decision-timeout/2026-10-08-a40-native/ (389f346): through system1-omni's frontend and its native Open-Jev worker (4a79980) with Open-Jev-9B. 5 s: the first decision (8.7 s) is cut, clean stop at 12.3 s. 30 s: both decisions complete, then the model answers BLOCKED. 1 s: clean failure, no hang.
• docs/results/decision-timeout/2026-10-07-a40/ (34cc19b): the same three runs through system1-omni's frontend in front of Open-Jev's reference server with Open-Jev-9B; plus the native worker's out-of-memory log for Open-Jev-27B-v1.1 on the 46 GB A40 and the measured export cost (37 s, 59 GiB RAM).
  Each folder has the records, the commits, the hardware and the scripts as run; see its README.

Verification: ruff and ty clean; 648 passed. One failure (test_decision_models_laya.py::TestFromEnv::test_an_agent_without_system_one_is_a_config_error_naming_the_method) fails the same on main on this machine.

[Birfy]

Review of current head 389f3463024d6e64683d7ec8486e738417873c38 against 9790b52ef4d78a2215e8dd5a526d3a8daea9875a.

I found no blocking code issue in this focused fix. JevDecisionsClient now rejects non-finite and non-positive explicit timeout_s values before constructing the httpx client, reports a configuration error naming timeout_s, and the regression test covers zero, negative, NaN and both infinities. The real-server deadline evidence is also recorded with the runs and environment details. Core 3.11, 3.13, Windows 3.11 and full CI pass; browser is skipped by configuration.

Disposition: no code findings from this review. This is a narrow bug fix with reproducible trace evidence; I do not consider a separate feature recording necessary here.

[Birfy]

Reviewed current head 5bfd358b80f09996d29b81e1d8f2dda9d66712d8 against base 5d89f85e702fc8dbdb7b719721e3cb0dda0bef4f.

The fix closes the explicit-timeout validation gap: JevDecisionsClient rejects non-finite and non-positive timeout_s values before creating the httpx client, and reports a configuration error naming the parameter. Regression coverage includes zero, negative, NaN and both infinities. Normal timeout/deadline behavior is unchanged. The real-server evidence records the 1 s, default 5 s and 30 s cases with run artifacts and environment details.

I found no blocking code issues. Core 3.11, core 3.13, Windows core and full CI pass; browser is skipped by configuration. Approved.