fix(reflect): enforce max_tokens after forced synthesis

[koriyoshi2041]

Problem

max_tokens is enforced with a post-hoc rewrite when reflect finishes through the done tool, but forced final synthesis returns its answer directly. Long or context-exhausted runs can therefore exceed the caller's visible-length target.

Closes #4156.

Fix

• Extract the existing final-answer rewrite into a shared helper.
• Apply it to both done and forced-synthesis completions.
• Run the rewrite before structured-output extraction so structured output is derived from the bounded answer.
• Include rewrite usage and trace data in forced-synthesis accounting.

Test

• uv run --project hindsight-api-slim pytest -q hindsight-api-slim/tests/test_reflect_agent.py -k 'not TestDirectiveLeakageOnEmptyBank and not TestContextOverflowIntegration and not TestMentalModelShortCircuitRealLLM' (55 passed)
• uv run --project hindsight-api-slim ruff check hindsight-api-slim/hindsight_api/engine/reflect/agent.py hindsight-api-slim/tests/test_reflect_agent.py
• uv run --project hindsight-api-slim ruff format --check hindsight-api-slim/hindsight_api/engine/reflect/agent.py hindsight-api-slim/tests/test_reflect_agent.py
• git diff --check

The four excluded tests require an embedded PostgreSQL extra or live LLM credentials; the first full-file run reached 55 passing unit tests before those environment-only setup errors.

Risk

Forced synthesis may make one additional LLM call when its answer exceeds max_tokens. Calls without max_tokens and answers already within budget are unchanged, and the rewrite remains governed by reflect_max_completion_tokens rather than using the visible target as a hard transport cap.

[koriyoshi2041]

Closing as superseded by #4180, which covers the same forced-synthesis length boundary and landed the stronger shared rewrite path. Thanks for picking this up.