fix(reflect): cap done tool answers

[Sanjays2402]

Fixes #2756.

The normal done tool path now applies the same capped rewrite used by direct-text completions when an answer exceeds max_tokens. The rewrite's token usage and trace entry are included in the final result.

Regression coverage confirms the done-tool answer is uncapped before this change and rewritten with the configured budget afterward.

[benfrank241]

Thanks @Sanjays2402 — and nice root-cause report on #2756.

I verified this locally: the fix faithfully mirrors the existing short-circuit rewrite path (same prompt, scope, and max_completion_tokens=max_tokens), plumbs max_tokens into _process_done_tool correctly, and threads final_usage + the final_rewrite trace entry through cleanly. test_reflect_agent.py passes 57/57 deterministic tests including the new test_done_tool_answer_respects_max_tokens (the 3 errors are just LLM API key required on the real-LLM classes, which run in the core-LLM CI job).

Routing to @nicoloboschi for the merge call, because this is a behavior decision as much as a bug fix. Today two of the three final-answer paths (forced-final, direct-text short-circuit) enforce the cap; the done path — the common one — does not. Applying the cap here means an extra LLM round-trip + a lossy summarization rewrite on the hot path, turning max_tokens into a hard, rewrite-enforced contract rather than a soft hint. It is scoped to mental-model refresh (the only place max_tokens is populated, per #1076), so it only fires where a user opted into a budget — but whether that budget should be hard-enforced via a second call vs. best-effort is a product call for you.

Fix looks correct and ready if you want it; flagging the design tradeoff so the decision is yours. cc @nicoloboschi