Sequence-first causal reconstruction for historical tool-use replay

Historical tool-use replay should treat event sequence as the primary order and timestamps as causal guards, not as an alternate sort key. Build each model input only through the target caller boundary; require complete ID/function-matched prior tool-call/result pairs; keep post-decision results outside the request; and quarantine sequence/timestamp inversions instead of repairing them silently. Initial caller turns are usually the strongest metadata-derived cases because they need no historical result reconstruction, but even they remain constructed prefixes unless an actual request body or digest is bound. This applies to reset-context voice nodes with captured prompt/tool snapshots; it does not prove the recorder captured every byte sent to the model.

A separate anchor rule is essential: if the prompt snapshot already contains the emitted assistant anchor and the turn log repeats it, suppress only an exact duplicate (or an authoritatively identified duplicate). Do not blanket-drop the prompt anchor merely because the first logged turn is agent speech.

Evidence/example — 2026-09-16: read-only comparison of the retained MC causal extractor, carrier source-native importer/renderer, reference Qwen role mapper/tokenizer, and strict baseline parser. The MC extractor already enforces caller-before-decision timing and filters observed results by both sequence and timestamp; the carrier path adds strict call/result linkage but lacks timestamp causality; the Qwen path explicitly maps developer to system and verifies nonthinking template/tool fidelity. A contact replay reuse plan records the combination and its remaining model-visible-result fidelity gap. No inference, training, fresh held-out content, or backend state was used.