feat(providers/anthropic): surface refusal stop_reason as content-filter

[jscottmiller]

Summary

Map Anthropic's refusal stop reason to a content-filter finish reason and
preserve the block details, so downstream callers can tell a safety block apart
from a normal stop and surface why a turn produced no content.

Anthropic's real-time safety classifiers end a response with
stop_reason: "refusal" and a stop_details object (category,
explanation), and emit no content. Previously mapFinishReason fell through
to FinishReasonUnknown and the details were dropped.

Changes

• mapFinishReason: "refusal" -> FinishReasonContentFilter.
• Parse stop_details from the message_delta raw JSON (the Anthropic SDK does
  not model it) and attach it to the finish stream part via a new
  RefusalMetadata{Category, Explanation} provider-metadata type
  (registered, marshal/unmarshal, GetRefusalMetadata getter).

Testing

• go test ./providers/anthropic/ (new refusal_test.go covers the finish-reason
  mapping, stop_details parsing, and metadata round-trip).

Notes

Consumed by coder/coder PR for chat "Response blocked" messaging.

---

Coder Agents generated.

[jscottmiller]

Verified the claim in the message_delta comment ("The Anthropic SDK does not model stop_details"): true for the SDK we pin, with one nuance.

• Our chain (charmbracelet/anthropic-sdk-go replaced by the coder/anthropic-sdk-go fork) is based on upstream v1.26.0; no stop_details/StopDetails anywhere in the module (it does model the refusal stop-reason enum).
• Upstream added structured stop_details in v1.29.0, and v1.46.1 fixed the beta accumulator dropping stop_details from message_delta events.

So the raw-JSON parsing is the right call today. Suggested follow-ups, not blocking:

1. Reword the comment to "the SDK fork we pin (v1.26.0 base) does not model stop_details" so it does not read as a claim about current upstream.
2. When the fork is rebased onto >= v1.29.0 (ideally >= v1.46.1), replace parseAnthropicRefusal with the typed field.

---

Coder Agents generated (on behalf of @jscottmiller).

[jscottmiller]

Merge-order note relative to #40 (the main -> coder_2_33 sync):

• A test merge (git merge-tree) of the two branches shows this PR adds exactly one conflicted file to the sync, providers/anthropic/anthropic.go; provider_options.go and the new test auto-merge. chore: merge main into coder_2_33 #40 already conflicts with current coder_2_33 in two OpenAI files (it was cut before fix(providers/anthropic): send thinking display #39 merged), so it needs a refresh regardless.
• Resolution is mechanical either way: upstream main keeps the same structure at all three insertion points (mapFinishReason switch, stream event switch, finish-part ProviderMetadata), and contains no refusal/stop_details handling of its own.
• chore: merge main into coder_2_33 #40 does not bump the coder/anthropic-sdk-go pin (still v1.26.0 base), so the raw-JSON stop_details parsing here remains necessary after the sync.

Since the refusal mapping is not Coder-specific, it is also a candidate to send upstream to charmbracelet/fantasy rather than carrying it as a fork patch long-term.

---

Coder Agents generated (on behalf of @jscottmiller).

[jscottmiller]

/coder-agents-review