[nicoloboschi]

Fixes #5612.

Problem

OpenRouter reports an upstream that died mid-generation as HTTP 200 with finish_reason: "error", no tool calls, and the cause on the choice. The tool-call path read choices[0] directly, so this became an empty tool-call result: reflect treated a transport failure as "the model chose no tool", spent its single loop-wide retry, then answered without the forced recall / search_observations step. Nothing recorded which upstream failed.

Fix

• openai-compatible: a finish_reason="error" choice raises a retryable ProviderResponseError (with the choice's error and the served upstream), checked in the helper every path reads its first choice through. call_with_tools now uses that helper and retries ProviderResponseError with backoff inside its existing attempt budget, like call() already did. call() also stops accepting partial text from an errored generation.
• litellm: LiteLLM maps "error" to "stop" but keeps native_finish_reason and error in the choice's provider_specific_fields; the same check reads them there, and both call paths retry it.
• tracing: spans carry hindsight.provider.served_by (OpenRouter's response provider), and every retried attempt is recorded as a failed span + success=False metric with its billed usage — previously it left only a log line.

Behaviour changes

• When every attempt errors, the tool call raises instead of returning an empty result (reflect's own capped LLM-error handling takes it from there).
• The tool path now also rejects an error payload / empty choices with a clear ProviderResponseError (empty choices used to be an IndexError).

Tests

• tests/test_openai_compatible_finish_reason_error.py: reporter-shaped responses — tool call retried then answered; raises naming the upstream when retries run out; call() rejects partial text; retried attempt traced with upstream + usage; OTel span attribute; LiteLLM (call and call_with_tools) through LiteLLM's real response conversion.
• Tool-path fakes in 8 test files now set error=None (the path never checked it before).

Not in scope (from the issue's ask 3): surfacing "a forced step was skipped" in reflect / refresh result metadata, and a per-step forced retry.