feat(responses): translate legacy <think> XML into native reasoning fields

[weselben]

Destination

A response-path feature in GoModel that detects old Anthropic-style
<think>...</think> (and known equivalents) embedded in model output
and rewrites them into each API surface's native structured reasoning
field, so clients get reasoning on the right field instead of an XML
blob inside the message content.

End state (one-line acceptance): a model that emits <think>x</think>hi
returns, on the OpenAI chat endpoint, { "message": { "content": "hi", "reasoning_content": "x" } }; on the OpenAI responses API, a
reasoning item with that text; on the Anthropic messages endpoint, a
content block of type: "thinking" carrying that text. Nothing about
requests changes. Streaming partial tags reassemble correctly across
chunks.

Notes

• Scope: response path only. Request side already has GoModel's
  compression / guardrails pipeline and is not in scope here.
• Three API surfaces: /chat/completions, /responses, /messages.
  Each has a different native reasoning field shape.
• Skill: consult writing-for-agents style for PR body, comments,
  commit messages; consult grilling and domain-modeling for any
  ambiguous decision.
• Worktree: .worktrees/feat-think-block-translation on branch
  feat/think-block-translation. PR target: weselben/GoModel.

Decisions so far

• Q1 — opt-in vs always-on default? — Default on, per-surface opt-out via env; WebUI per-surface override is a separate ticket (T8).
• Q2 — exact XML tag set to recognize — Research pass on known legacy tag variants (T9), then a configurable comma-separated tag-pairs list (T10). Regex-based matching rejected for chunk-boundary safety.
• Q3 — Anthropic thinking-block signature policy — Three-way enum (off/unsigned/redacted) on the messages endpoint (T11). Off is the default.
• Q4 — chat completions reasoning field name — Ships with reasoning_content as the repo-wide canonical field. Client-capability detection is a relabeled research ticket (T38).
• R5 — streaming partial-tag state-machine precedent — Dedicated State type in internal/thinkextract, not a reuse of streaming.StreamBuffer.
• Q6 — hook point: per-provider vs post-format — Shipped as a single canonical-chat hook at the orchestrator (chat + messages + responses-via-chat). Awaiting user feedback on whether to refactor to per-provider.
• T7 — native Responses API stream transformer — Open task, depends on PR feat(responses): translate legacy <think> blocks into native reasoning field #41 merge.
• T8 — per-surface opt-out — Open task, depends on PR feat(responses): translate legacy <think> blocks into native reasoning field #41 merge.
• T9 — research known legacy reasoning tag variants — Open research ticket.
• T10 — tag-pairs env + config — Open task, blocked by T9.
• T11 — /v1/messages thinking-block policy — Open task, depends on PR feat(responses): translate legacy <think> blocks into native reasoning field #41 merge.

Not yet specified

• 99%-coverage test scope boundaries (draft PR is at 91.8% package coverage)
• Dashboard runtime-settings pattern for per-surface override (referenced by T8)
• Responses reasoning.summary population — deferred
• Per-provider vs canonical-chat hook — awaiting user feedback on Q6

Out of scope

• Request-side <think> translation (clients send native fields
  already; this PR is response-side only).
• Reasoning-effort / model-side reasoning controls.
• Compression / guardrail interactions (separate pipeline).
• Responses API reasoning.summary field population beyond raw
  reasoning text — left as a follow-up if/when needed.
• Fast-path chat passthrough and the native Anthropic messages
  passthrough (documented exclusions).

[weselben]

note: kimi - session_ee879c1e-ce72-4a7e-8af3-6376ec845892