fix(anthropic): degrade thinking replay before the turn is rejected

[probepark]

What

Detect locally that the latest assistant turn cannot replay its thinking block, and degrade the replay during the first build — instead of sending the request, being rejected, and repairing afterwards.

Why

Anthropic opens and closes a thinking block with no thinking_delta and no signature_delta when it withholds reasoning. The block lands in history with empty text and no signature, convertAnthropicMessages drops it, and the assistant turn goes back carrying only its tool_use. Anthropic validates the latest assistant message against the turn it produced and rejects the missing block:

messages.N.content.M: `thinking` or `redacted_thinking` blocks in the latest assistant
message cannot be modified. These blocks must remain as they were in the original response.

Captured from a live rejected request (headroom HEADROOM_DEBUG_DUMP=full):

• 1214 messages, model claude-opus-5, stream=true
• 29 signed-but-empty thinking blocks forwarded intact
• latest assistant message: ['tool_use(bash)'] — no thinking block
• the proxy's pre-compression snapshot is byte-identical for that turn, so all 29 empty blocks came from the client and the proxy emptied none

The rejection was already recoverable through the existing replay repair, but only after a full round trip. This makes the degradation happen before the request is sent.

Scope is deliberately the latest assistant turn only, because that is the turn Anthropic validates against its own output; an older hollow block must not cost the rest of the replay.

Testing

New packages/ai/test/anthropic-unreplayable-thinking.test.ts:

• hollow latest turn -> no native thinking replayed (before fix: Expected 0, Received 1)
• healthy signed latest turn -> both thinking blocks preserved
• hollow block on an earlier turn -> latest turn's thinking preserved

Also:

• bun test packages/ai/test: 2293 pass / 337 skip / 4 fail — the same 4 failures a clean checkout has (buildAnthropicAuthConfig base URL, 3x openai-responses cache affinity), confirmed by stashing this diff
• bun --cwd=packages/ai run check: biome + tsc clean
• artifacts/issue-3670-anthropic-cache-eval.json regenerated (it pins the sha256 of providers/anthropic.ts)

Related: #4159 stops the same rejection from recurring every turn; this one stops it from being paid at all.

GJC verdict

gajae.pr-review-verdict.v1 needs-human sha256:b8c00d85b50676cd1032fbdbe210d094dfdef5c1 reviewer:human evidence:local bun test packages/ai/test + bun --cwd=packages/ai run check

---

• Target branch dev
• bun check passes
• Tested locally
• CHANGELOG updated (if user-facing)
• Verdict matches exact PR head, not earlier commit