fix(compact): count and shed reasoning_content

[zhanghanduo]

Summary

With thinking_format: reasoning_content and thinking_in_history: true, previous assistant turns resend their reasoning, but compaction previously ignored it. A llm-hub deepseek-flash TBS run estimated about 3.5k tokens for a 66–80k-token prompt, preventing the tiered compactor from recognizing sufficient relief.

This PR:

• Counts reasoning_content in loop, state-aware middleware, and portable SummarizationMiddleware token estimates.
• Adds strip_stale_reasoning(messages, keep_last) without mutating the source history.
• Retains at least the latest assistant reasoning in Tier 1 (keep_tool_result=-1 still disables Tier 1 filtering). Tool-result compression candidates and Tier 2 input retain the latest assistant reasoning.
• Updates the change fragment and documents the keep_tool_result=0 behavior.

Proxy validation

A real llm-hub deepseek-flash three-turn tool-call chain was replayed with reasoning on all turns, only the latest turn, and no turns. All three requests returned HTTP 200 and the correct answer (1296). Prompt tokens were 599, 487, and 479 respectively, confirming the field contributes to the prompt and older reasoning can be omitted for this proxy.

This does not establish compatibility with the self-hosted DSV4 checkpoint, which was unavailable, or answer-quality effects on long tasks.

Testing

• New reasoning-specific tests cover all three estimate paths, both portable middleware estimate branches, threshold-triggered summarization, Tier 1 retention including keep_tool_result=0 and -1, tool-result compression, and non-mutation.
• Local full suite: 2335 passed, 1 skipped.
• Ruff and Pyright pass.

[zhanghanduo]

Proxy check: stripped historical reasoning is accepted (llm-hub deepseek-flash)

I tested the PR's assumption with a real request. First I ran a 3-turn tool-call chain ((17×23) → +41 → ×3) where every assistant turn returned reasoning_content. Then I sent the same history three ways:

Variant	HTTP	prompt_tokens	Answer
Reasoning on all 3 turns	200	599	1296
Latest turn only (this PR)	200	487	1296
Stripped entirely	200	479	1296

• The proxy doesn't require reasoning_content on earlier turns. Even stripping it entirely gives no error and the correct answer.
• The drop in prompt_tokens confirms that reasoning sent in history is counted toward the prompt, which is the undercount this PR fixes.

Not covered: the self-hosted DSV4 checkpoint (tbs-1004-v1-e3, service currently down) and any effect on quality in long tasks. A 3-turn arithmetic chain can't show either.