The estimator has no seam: EKF hardcoded in the loop, EKFParams welded into run_simulation, its name baked into the traces — a UKF touches 6–10 files

[dim-sous]

The probe

Swapping EKF for UKF is v6 on the roadmap and the cleanest test of the platform's one rule: a change in A must not require a change in B. It fails today.

What is welded, with evidence

coupling	where	what a swap forces
Hardcoded class	core/simulator/core.py:44 import, :82 ekf = ExtendedKalmanFilter(bp, tp, ekf_p, thp, elp)	edit the simulator core
Config type in the signature	core/simulator/core.py:65 ekf_p: EKFParams — one of 15 positional params	edit run_simulation and its 7 call sites
Constructor takes 5 param objects	core/estimators/ekf.py:47	every estimator must accept the same five, or the core changes again
Algorithm name in the trace schema	traces.soc_ekf / soh_ekf / temp_ekf / vrc1_ekf (core/simulator/traces.py:53-56), record_mpc(ekf_state=…)	ledger, plots and traces say "ekf" for a UKF — or rename in 4 files
Comments and warnings	core/mpc/economic.py:70,263,296, core/planners/stochastic_ems.py:125, strategies/ems_economic_mpc/strategy.py:5	cosmetic, but drift compounds

Files touched for a UKF today: 6 minimum, ~10 honestly. For a change that should be one new file plus one line naming it.

What already works — the contract exists, unstated

The loop reaches the estimator through exactly two calls: ekf.get_estimate() (core.py:96) and ekf.step(u, y_meas) (core.py:164). A UKF exposing those two is a drop-in at the call site. The port is there; nothing declares it.

The pattern is one seam over

core/simulator/strategy.py already does this right: Strategy is a frozen dataclass naming a planner and an optional MPC through _PlannerLike / _MPCLike Protocols; the loop calls strategy.planner.solve() and never branches. v5b was added as a new folder with zero core edits, and tests/test_greek_milp_bidding_e2e.py::test_v5_strategy_not_affected_by_phase3 proves the old traces stayed bit-identical. The estimator is exactly what did not get that treatment.

Done when

• An EstimatorLike Protocol (step, get_estimate) owned by the core, next to _PlannerLike.
• The concrete estimator chosen at the composition root — Strategy.estimator or a registry entry — never inside the loop.
• Estimator config carried by the estimator, not by run_simulation's signature.
• Trace fields named for the quantity (soc_est), not the algorithm.
• A registry with provenance (RCWS A run does not describe itself: results/ holds 27 files named by version, with no config, seed, conditions or git stamp — two runs of one version overwrite each other #6: "a new tuning is added by config; a new model plugs in through the one interface").
• Proof: EKF traces bit-identical before and after. Then v6 is a new file.

Phase 1 of the platform epic. Depends on the layout issue (#1).