server: per-request reasoning_budget_message and exact cache round-trip with a wrap-up message

[pugant]

Follow-up to #69 — completes the reasoning-budget round-trip when a wrap-up message is configured. Two commits, both running in production since Aug 16 on Qwen3.8-27B (hybrid recurrent) + MTP n-6.

1. reasoning_budget_message per-request on the OAI chat path

The generic copy-through of remaining body keys in oaicompat_chat_params_parse (for (const auto & item : body.items())) skips keys already present in llama_params, so the CLI value pre-populating reasoning_budget_message silently shadowed any per-request one — the body field was ignored on /v1/chat/completions (it worked on /completion, which reads it directly). Read the body field explicitly with the CLI value as default, mirroring what the thinking_token_budget alias handling does for the budget count.

2. Separate newline between wrap-up message and end tag

Chat templates re-render a resent assistant turn as '<think>\n' + reasoning_content|trim + '\n</think>\n\n' (verified against the Qwen3.8 chat template extracted from the GGUF). The |trim strips the trailing newline from the extracted reasoning, so the forced tail '\n' + message + end_tag from #69 diverges at the message/end-tag boundary: the re-render adds a '\n' the generation never emitted, and the resend pays a checkpoint rollback instead of an exact hit.

Fix: emit '\n' + message + '\n' + end_tag when a wrap-up message is set; keep the bare '\n' + end_tag when empty — that is the exact-hit case already verified in #69 and it stays untouched.

Verification (two scripted runs on the production image)

• Run 1, wiring only (commit 1): the per-request message reaches the forced tail and appears in the extracted reasoning; but a resent budget-truncated turn paid a checkpoint rollback (lcp=280 on cached=534; lcp=1338 on 1398) — divergence exactly at the message→end-tag boundary.
• Run 2, both commits: regression check with an empty message and wrap-up round-trip with the message set: zero cold fallbacks, zero rollbacks — exact cache hit with the message in the forced tail; multi-turn session with two truncations also clean.
• Production since Aug 16: 9/10 dense agent tasks hit budget exhausted → forcing → forced complete with the per-request message active, all clean (27/27 marker-matched runs).

Note

Default remains no message (--reasoning-budget-message, default: none). An explicitly configured message should not carry trailing whitespace: the template |trim would strip it and the resend would pay a bounded rollback (covered by the checkpoint salvage from #69).

🤖 Generated with Claude Code