fix(prefill): inject immutable retained-token cap

[FluffyAIcode]

Summary

• inject sink + window directly into PrefillJobStore as max_retained_tokens
• use the explicit cap for both submission preflight and runtime reservation
• retain compatibility-derived capping only as a fallback for generic callers
• prevent long prompts from reverting to full-token snapshot estimates when a runtime compatibility field is absent

Root cause

The active worker used the correct source and launch arguments, but its runtime capacity path observed a compatibility object with window_size=0. It therefore estimated 2731 * 400000 = 1.0924 GB and rejected the snapshot after a long Prefill. Relying on a serialized compatibility field made the same failure recur after unrelated changes.

The explicit deployed cap is sink + window = 2052, yielding a stable 820.8 MB estimate.

Tests

• explicit cap overrides a missing compatibility window
• worker startup injects configured sink/window
• existing preflight and reservation behavior
• targeted suite — 19 passed
• full platform-neutral suite — 885 passed, 1 skipped
• Python compile and diff checks pass

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.