grill-perf: workload deadline ceiling (total_ms ≤ 600000) rejects legitimate long-output cells

[plotarmordev]

Workload::validate rejects total_ms > 600_000 with invalid deadline or buffering limits. A 400-token thinking-on prose cell on a 2×GB10 GLM serve runs ~20 tok/s per stream; with concurrency 2 and DFlash2 deferral one wave already took 112 s, and a long-context decode cell can legitimately exceed 10 min. Suggest raising the cap (or making it a CLI override) and improving the error to say which limit failed. Evidence: run dir decode-main-3584-attempt1-idle-timeout on our kit, idle_timeout after 120 s with Waiting: 2, Deferred: 2 on the server.

[plotarmordev]

Resolved in PR #20: an explicitly declared finite total deadline may now be up to one hour, with field-specific deadline/buffer admission errors. The selected workload limits remain recorded and identity-bound; no hidden override or retry was added. Important distinction: raising total alone does not fix an idle timeout. Silent prefill/queueing needs an appropriate explicit idle budget, still <= total. CLI tests cover admission and separate total/idle expiry; no hour-long model test was performed.