On agentic runs, get_session re-tokenizes the full trajectory to compute tito_session_mismatch on every record fetch, synchronously on the session-server event loop. At 64k-context SWE scale this blocks record retrieval — the client-side collect_records GET times out (120s), returns empty, and the sample gets marked ABORTED. On a recent run that was ~19k timeouts, ~58% of samples per step, which cascaded into whole groups being dropped by the infra-failure filter.

The mismatch is a diagnostic canary for chat-template / tokenizer drift, logged as tito_session_mismatch_rate. It is not a training-correctness guard — training tokens come straight from the inference response (prompt_token_ids / output_token_logprobs), and the per-turn alignment asserts on the training path already enforce that. So computing it on every session is pure overhead once the template is stable.

This gates it behind --tito-session-mismatch-sample-rate (default 0.0, off). Set 0.02 for a cheap drift signal, 1.0 to restore the previous always-on behaviour. Sampling is a deterministic crc32(session_id) so there is no RNG on the hot path.