Explainability is a question about structure, not about magnitude

[amburger66]

Stacked on #143 — review that one first; this diff is only the last commit.

The problem

The trimmer drops a recording no parameter value can explain, so a chaotic replay cannot drag the fit to a compromise nobody supports. It decides with an RMS bar: trim_rms_factor * noise_sigma = 0.1.

Under interval scoring that bar is applied to two residuals that are nothing alike:

• a missing-cascade penalty — the twin failed to reproduce which dominoes fall
• a timing term — both sides saw a domino fall, at different moments

On run_20260820_141450 one penalty was 2884 in SSE while the entire timing disagreement was 5.3. So the bar is really a penalty detector with a wildly miscalibrated tail, and it rejects structurally perfect data:

per-trajectory best RMS ['inf', 'inf', '1.21'] vs threshold 0.1000
NO trajectory is explainable at any candidate params

That 1.21 is a theta at which the twin reproduces all four falls with zero penalties — the same theta the sweep reports as explaining the data 1084x better than baseline (0.2402 -> 5.321 against 5769 at the agent's belief). The fit refused to run on the best explanation it had, because the timing error it exists to reduce was used as grounds for refusing to fit.

The change

Three outcomes rather than one number:

at the best candidate	verdict
nothing measured at all	not explainable — nothing to have explained
any missing-cascade penalty	not explainable — the wrong dominoes fall
penalties absent, timing only	explainable, and the fit does its job

The first row is not padding: a cascade-free segment emits no penalties because it emits nothing, and reading that as success would reinstate the inversion that made an empty residual vector an RMS of 0.0.

Penalties are counted from the id sets — a penalty is a domino one side has and the other does not, i.e. the symmetric difference — never by recognising penalty-sized numbers in the output.

The per-step objective is untouched. There the residuals really are a dimensionless fraction of typical motion and 0.1 means what it says, so the RMS bar still decides; gated on score_observed_only.

Worth a reviewer's attention

• This is more permissive. A recording where the twin gets the right dominoes falling but badly wrong timing is now kept and drives the fit. I believe that is the data the fit should use, but it is a behaviour change, not a refactor.
• reproduced_cascade asks whether any candidate reproduced the cascade, not whether the RMS-best one did. Those differ, and "could this model ever produce this outcome" is a question about the whole grid.
• min_explainable_fits gains a third return and the explainability cache a third element.

Verification

177 tests pass, mypy clean over code_sim_learning/, pylint 10.00/10. The new test fails under a revert of the structural switch — and exactly that one test fails, since it passes its own SysIdConfig rather than touching CFG, so interval scoring does not leak into the rest of the module.

Not yet run end to end: the replay (exp_domino_real_replay.yaml, added in #143) is the cheap way to see whether this turns the pinned 0.24 into an actual fit.