Emit Anthropic thinking blocks from reasoning channel

[HXYerror]

Part of #1. Pairs with #6 (chat-completions side) and #7 (responses side).

Goal

Today the Anthropic adapter actively suppresses thinking block output. With reasoning models live (gpt-5, o3, codex…), Anthropic clients (Claude Code, Anthropic SDK) lose all visibility into the model's chain of thought. We should emit thinking blocks whenever the upstream provides reasoning material.

Current state

src/routes/messages/non-stream-translation.ts:305 says verbatim:

// GitHub Copilot doesn't generate thinking blocks, so we don't include them in responses

That comment is stale as of the GPT-5 / codex rollout. Additionally the thinking_delta and signature_delta types declared in src/routes/messages/anthropic-types.ts:145–146 are dead code — no producer in the repo.

On the input side, thinking blocks are flattened into plain text:

• src/routes/messages/anthropic-types.ts:21–24 — thinking? payload field is silently ignored
• src/routes/messages/non-stream-translation.ts:146–154, :195–211 — inbound thinking content blocks → text

Tasks

• Remove the suppression in non-stream-translation.ts:305 and emit { type: 'thinking', thinking: '...', signature: '...' } blocks when:
  • Chat-completions path: message.reasoning_content is non-empty (depends on Preserve encrypted_content for multi-turn reasoning #6)
  • Responses path: any output[] item with type: 'reasoning' (depends on Add upstream Responses API service client #3)
• In stream translation, wire thinking_delta and signature_delta events. The Claude streaming spec orders them: content_block_start (type=thinking) → repeated thinking_delta → final signature_delta (carries the opaque CoT signature) → content_block_stop.
• Use signature to carry encrypted_content for the Responses path (Anthropic's signature field is the natural home for an opaque verifier blob; see Model-to-endpoint routing (chat vs responses) #5)
• On the input side, when an Anthropic thinking block with signature is included in history, don't flatten to text — preserve it as a reasoning item that gets forwarded upstream (cross-link Model-to-endpoint routing (chat vs responses) #5, Reasoning types & reasoning_effort passthrough in chat-completions #7)

Acceptance criteria

• Claude Code shows the model's thinking inline when running against gpt-5 / o3 / gpt-5.3-codex
• Two-turn flow round-trips thinking.signature byte-exact (no upstream verification errors)
• No regression on tests in tests/anthropic-response.test.ts

File pointers

• src/routes/messages/non-stream-translation.ts:305 — remove suppression
• src/routes/messages/anthropic-types.ts:21–24, 145–146
• src/routes/messages/stream-translation.ts — state machine

[HXYerror]

Duplicate of the corresponding #9 — closing to avoid confusion.