Defer to guessed starting values no more than the prior does, and report the belief honestly

[yichao-liang]

Why

Seed 0 of the Domino round fixes_r3 lost its test level after it stopped trusting its fitted belief, for two reasons this PR addresses (#219 fixed a third, refits drifting with the last fit's answer).

1. The fit deferred to the agent's guesses beyond its prior.
   The grid sweep counts a candidate within 5% of the model-bias SSE as data-equivalent and keeps the one nearest the anchor, and the anchor ablation pins moved parameters back to their anchors.
   That suits calibrated baselines (a supplied domain base).
   Under code_sim_learning_prior_spans_bounds the anchors are the agent's guesses: on seed 0's level-1 recording the guessed spinning friction 0.005 stayed put although its best candidate, 0.67 with the world at 0.5, fit 2.8 nats better (5.6 likelihood floors).
2. The report contradicted itself.
   It printed "most likely " next to the central 68% interval of the belief factor, which can exclude both the point estimate and a mode at a bound: seed 0 read "most likely 0.05; 68% posterior interval [0.2694, 0.291]".

What changes

• SysIdConfig.anchors_are_guesses (from code_sim_learning_prior_spans_bounds): the flat band drops its relative term (flat_band_frac), so data-equivalence is the likelihood floor the belief scores with, and the anchor ablation does not run. Supplied-base arms keep both.
• The belief report gives each factor's mode and its 68% highest-density interval, which contains the mode, and names the fit's point estimate when it lies outside that interval; the note under the report says the planning model runs at the point estimate.
• test_run_rollout_sysid_fit_cache_and_report_isolation starts from the default config: test_continual_joint_belief leaves belief_joint_draws set, which sent the fit down the joint-belief path whenever the two tests ran in that order.

Evidence

The scripted from-assets harness fits seed 0's level-1 recording (7 segments) with and without the change; the world's values are known.

• Fit time: 230 s with the change, 682 to 907 s without (2,403 s before Make the rollout fit about 3 times faster without changing its results #218); the anchor ablation, 6,293 of 7,735 replays, no longer runs.
• Replay error: 7.139 with the change, 7.187 without.
• Domino lateral friction 0.388 (world 0.5, was 0.954), spinning friction 0.165 (0.5, was 0.005), rolling friction 0.0067 (0.006, was 0.0005); table rolling friction 0.0010 (0.001, was 0.0025).
• Domino restitution moved away from the world: 0.245 (world 0.02, was the guessed 0.05).

Tests

• New: a calibrated anchor keeps its value when a better candidate lies inside the relative band, a guessed one does not; the ablation runs for calibrated anchors only; the highest-density interval holds a mode at a bound that the central interval excludes; the report names a point estimate outside the belief.
• Local CI replay: the static checks and the 8 shards.

🤖 Generated with Claude Code