[Benchmark] Prevent Pilot overlays from regressing dispatched-action history

[enkerewpo]

Problem

The frozen EmbodiedBench V12 run replaced system/pilot/src/planner.rs with an older benchmark overlay. That overlay removed the dispatched-RTDL history retained by commit 583d1e0 and added a generic "unchanged successful action" rejection instead.

In the preliminary first-12 run, five action-bearing Robonix cells ended one turn with I stopped an unchanged failed-action loop..., then the corrective submission repeated a navigation that had already succeeded. This is an evaluation-profile regression, not evidence that current dev still lacks the dispatched-action fix.

Why it matters

A benchmark overlay must not silently remove mainline safety/state-continuity behavior. Otherwise the benchmark measures the overlay rather than the branch it claims to evaluate.

Acceptance criteria

• Record the exact Robonix commit and a machine-readable patch/overlay digest in every result.
• Fail preflight when an overlay removes the dispatched plan/call history contract or changes duplicate-action semantics without an explicit protocol revision.
• Replace the generic "same successful navigation" rejection with plan/call provenance and ordered feedback handling.
• Add a regression scenario where an older plan fails after a newer plan succeeds; the next corrective turn must retain both outcomes and must not repeat the successful physical action.
• Rerun the affected cells under a new protocol version; do not overwrite V12 evidence.

The cited first-12 results are preliminary diagnostics. Robonix is under active optimization, and these results do not represent final system performance.

[enkerewpo]

Closing this as incorrectly scoped. The duplicate-navigation behavior came from my local, unversioned benchmark overlay replacing current dev Pilot code. Current dev already retains dispatched RTDL action history in commit 583d1e0; this is not a Robonix repository defect. I will track it only in the benchmark protocol/evidence and will not attribute those affected cells to current dev.