concurrency-batch-all-land@fault: flaky  # New scenario; inherited from the raindrop-ai Python base SDK: whether concurrent calls land in one events/track batch depends on background-thread timing; under investigation with DEV-1143.
retry-429-hot-path@fault: flaky  # Inherited from the raindrop-ai Python base SDK: the forward() hot path only enqueues (well under max_ms), but the 429 retry is lost because flush() is not durable-on-return (CONTRACT 1.2) and process exit kills the pending retry (DEV-1143); it passes when the retry happens to land before exit.
size-limit-oversized-attachment@fault: missing_feature  # New scenario; the wrapper's capture path has no attachment concept, so the driver cannot map the step (exit 4 unmappable).
track-attachments@fault: missing_feature  # A DSPy forward() carries no attachment channel; the driver refuses attachments loudly (exit 4) rather than silently dropping them.
track-attachments@prod: missing_feature  # A DSPy forward() carries no attachment channel; the driver refuses attachments loudly (exit 4) rather than silently dropping them.
wrap-capture-usage@fault: missing_feature  # DSPy derives token usage from the LM response, not from a caller-supplied usage arg (forward() has no usage channel); the offline DummyLM reports no tokens, so ai.usage.* never populate.
wrap-traces-model-spend@fault: missing_feature  # DSPy model ids name the provider as a prefix (openai/gpt-4o) and the model span takes gen_ai.provider.name from it; the scenario's bare fixture model id has none, so the span carries the usage and models but no provider.
