fix: handle Anthropic 'refusal' stop reason instead of retry-looping

[jeffbryner]

Fixes #3624

Problem

When Claude Fable's safety classifiers decline a request, the Anthropic API returns HTTP 200 with stop_reason: "refusal" in the message_delta SSE event. Forge's StopReason enum didn't know that variant, so:

1. serde failed to parse the message_delta event (its stop_reason field is required),
2. the event fell into the untagged EventData::Unknown catch-all and was silently converted to an empty assistant message — the finish reason was lost,
3. the stream aggregator saw empty content + no tool calls + no finish reason and returned the retryable EmptyCompletion error,
4. the retry loop resent the identical request to the same model. A refusal is deterministic, so every attempt produced the same result:

ERROR: Empty completion received - no content, tool calls, or valid finish reason

repeated until max retry attempts were exhausted.

Fix

• forge_app (Anthropic DTO): add StopReason::Refusal, mapped to the existing FinishReason::ContentFilter (same mapping Bedrock uses for ContentFiltered and Google for SAFETY).
• forge_domain (stream aggregator): when the finish reason is ContentFilter with no tool calls, return a new non-retryable Error::Refusal with an actionable message ("rephrase the request or switch to a different model") instead of entering the retry loop.

Intentional behavior change beyond the reported bug

Because the check lives in the provider-agnostic aggregator, content-filter finishes from Google (SAFETY/RECITATION), Bedrock (ContentFiltered), and OpenAI-compatible (content_filter) providers now also end the turn with the clear error. Previously they re-entered the orchestrator loop (the turn only completes on Stop), appending an empty assistant message and re-requesting. In the mid-stream case the partial output (already streamed to the UI) is no longer persisted to conversation context — per Anthropic's guidance, partial output before a refusal should be discarded. Happy to gate this to empty completions only if maintainers prefer the narrower change.

Out of scope / possible follow-ups

• Automatic fallback-model retry on refusal (Anthropic's recommended recovery) — forge has no per-request model-fallback machinery today; this would be a feature PR.
• pause_turn and model_context_window_exceeded stop reasons are also missing from the enum and currently swallowed the same way; pause_turn needs orchestrator resume semantics, so it was deliberately excluded here.
• A #[serde(other)] catch-all on StopReason would prevent any future unknown stop reason from being silently dropped, at the cost of choosing a default mapping.

Testing

• New: wire-level SSE parse test for stop_reason: "refusal" (asserts it parses as a known event, not EventData::Unknown), mapping test, non-retryable error downcast test, and a mid-stream refusal-after-partial-content test.
• cargo test -p forge_app -p forge_domain -p forge_repo green (322 passed in forge_app), cargo clippy -D warnings clean, cargo fmt applied.

🤖 Generated with Claude Code