feat: carry reasoning across turns with per-backend replay policy

[AngeloDanducci]

Pull Request

Issue

Fixes #1201

Description

Carry reasoning across turns with per-backend replay policy for better thinking handling.

The policy is currently provider-agnostic (provider is unused, documented as a reserved hook). The divergent contracts from #1201 table — DeepSeek hard-400 without tools, Anthropic signed blocks — are not implemented here.

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code was added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

Attribution

• AI coding assistants used

---

Adding a new component, requirement, sampling strategy, or tool?

If your PR adds or modifies one of the types below, check the matching box. A checklist of type-specific review items will be posted as a comment.

• Component
• Requirement
• Sampling Strategy
• Tool

NOTE: Please ensure you have an issue that has been acknowledged by a core contributor and routed you to open a pull request against this repository. Otherwise, please open an issue before continuing with this pull request.

[ajbozarth]

Reviewed the full diff and traced the capture → _parse → policy → serialize path. Solid, conservative change — no correctness blockers. All 72 new/changed tests pass locally. Action items below.

Before merge

• Complete the live-model verification. The e2e test (test_reasoning_replay_e2e.py) is env-gated and CI-skipped, so nothing in CI exercises reasoning_content/thinking against a real endpoint. The PR body's own "still needs verification" note stands — this is the one thing that would catch a provider rejecting the injected wire key.
• Tighten the PR body to say this lands the consensus rule only. The title says "per-backend replay policy" and provider is a parameter, but the policy is currently provider-agnostic (provider is unused, documented as a reserved hook). The divergent contracts from #1201's table — DeepSeek hard-400 without tools, Anthropic signed blocks — are not implemented here. State that explicitly so reviewers/users don't assume DeepSeek/Anthropic are covered.

Non-blocking

• HF chat path is silently uncovered (see inline comment). Functionally fine given HF's capture gap, but a one-liner would prevent a future reader assuming coverage.

Verified separately

• should_replay_reasoning returns exactly len(messages) flags, so the zip in each backend never truncates.
• The adjacency check correctly maps to "this turn issued a tool call" — _acall_tools appends ToolMessage(role="tool") right after the tool-issuing turn.
• Default replay_reasoning=False preserves historical behavior.
• Real OpenAI Chat Completions never populates thinking, so the non-standard key is never sent there.

[AngeloDanducci]

Complete the live-model verification. - Live model verification was completed, just forgot to remove the line from the PR description when converting from draft.

[ajbozarth]

This LGTM as far as Claude reviews go , but I'd prefer to defer final approval to @jakelorocco after he returns since this is a decently sized change