Act on a from-assets test level only with a fitted model

[yichao-liang]

Why

Seed 0 of the Domino round fixes_r2 found its level-1 sim.fit() slow and uninformative (39 minutes for a 0.8% lower error), tuned its parameter values by hand from roll curves, and played the test level with an unfitted model.
The belief that chose its plan was therefore the prior around those hand-written values, not anything the recordings support.
The test-level gate (continual_require_model_on_test) only asked for a loadable model with RESIDUAL_FEATURES.
#218 makes that fit about 3 times faster (from about 40 minutes to about 12 on seed 0's recording) with identical results, so asking for a fit after each edit is affordable.

What changes

• _model_readiness asks the arm through _fit_readiness once a model with parameters loads.
• EMPIRIC from assets refuses skills_invoke and skills_execute_plan on a test level until sim.fit() has run on the current content of simulator.py; every edit needs a new fit.
  A fit that could not run (nothing explainable) still counts: the agent asked, and the belief then stays the prior.
• The supplied-base arms leave fitting to the agent as before, and real-to-sim fits nothing.
• The system prompt's gate section says which rule applies (model_gate_fit).

Tests

• The from-assets play test: the gate refuses before the fit and names sim.fit(), opens after the fit, and refuses again after an edit; real-to-sim needs only a loadable model; each arm's system prompt states its gate.
• Local CI replay: the static checks and the 8 shards (with Resolve a registry name to its own class, never a subclass that inherits it #213 applied).

🤖 Generated with Claude Code