Keep the from-assets arm's untested physics uncertain

[yichao-liang]

Stacked on #210 (from-assets-prompt-fixes); review that one first.

Why

Seed 1 of the Domino fixes_r1 round (EMPIRIC from assets) lost its test level.
Replaying its recorded test episode in the real environment shows the failing push needs the world's spinning friction (0.5) and rolling friction (0.006) on the dominoes, together with the gripper's follow-through touching the first bridge.
The agent never declared either friction, so its model ran both at PyBullet's 0 and predicted a cascade in every rehearsal.
Its level-1 recording, a straight chain, could not identify either value.
Declaring them would not have helped yet: the fit kept a parameter the data did not constrain at its starting value with a narrow belief, because its prior is 0.75 times the starting value.
A spinning friction started at 0.05 came out with a 68% interval of 0.057 to 0.12.

What changes

• A prior that spans the declared range.
  Under code_sim_learning_prior_spans_bounds, no parameter's prior is narrower than its declared range in fit space, so a parameter the recordings do not constrain keeps that whole range in the belief.
  The from_assets_opus menu entry sets it, and the arm refuses to start without it.
  The supplied-base arms keep the anchored prior, since their starting values are the base's calibrated defaults.
• Sampled undeclared materials.
  SceneBase.sampled_material_specs offers every engine material the scene does not declare as a parameter over a plausible range around the scene's own value (SAMPLED_MATERIALS: spinning friction 0 to 1, rolling friction 0 to 0.02, mass within a factor of 3, ...).
  The from-assets arm joins these prior-only factors to its belief (join_beliefs), so joint draws and physics sweeps vary them.
  They are never fitted and cost no replays.
  A rollout world applies a draw's value to the whole group; a fit that pins the material to its default gives each body its own value back, so a group the scene made different (the ground and the tables) stays different.
• Prompts.
  The from-assets system prompt and scene contract say that undeclared materials are sampled, how to fit or fix one, and that a declared parameter's prior spans its declared range.
  docs/amps/empiric-from-assets.md records the design.

Evidence

A scripted run of the from-assets harness on seed 1's round: level 1 is seed 1's own simulator.py and recorded actions.
On level 2 it fits, then rehearses two plans from the level's start on 16 joint draws each: seed 1's chain (3 blues, the 46-degree first hit, lost in the world) and the oracle-dynamics arm's arc (2 blues, won in the world).

	parent branch	this branch
Fit on the same data	SSE 3.245 to 2.285, support friction 0.80 to 0.30	same
Spinning and rolling friction in the draws	absent (PyBullet's 0)	sampled
Seed 1's chain, joint draws solved	12/16	10/16
Oracle arc, joint draws solved	11/16	9/16
Physics sweep of seed 1's chain	11/11 pass	21/23 pass; fails at spinning friction 0.97 and rolling friction 0.019

The sweep now names the two materials the chain hangs on.
The draws do not yet separate the two plans: the agent's own scene still predicts that seed 1's chain cascades at the world's material values, where the world's chain slides.
Real-environment replays rule out the materials, friction anchors, continuous collision detection, damping and the dominoes' massless top link as that difference.
The likely remaining cause is the push: the model replans it from its noisy estimate of the scene, and whether the gripper's follow-through catches a bridge 8.7 cm away decides the outcome.

Domino round fixes_r2 ran this head on 2 seeds and solved 1 of 2.
Seed 0 lost its test level because it declared both frictions, with ranges that exclude the world's values (spinning friction 0 to 0.05, rolling friction 0 to 0.002; the world has 0.5 and 0.006).
Neither change here reaches a declared range, so its 16 joint draws all cascaded.
A real-environment replay of its episode stalls at the world's values and cascades with either friction anywhere inside its ranges.
Covering declared materials whose range the recordings do not test is a follow-up.

The first acceptance run also caught a bug in the first version of the sampling: pinning a material's default homogenized the support group to the ground plane's friction, and the fit started at three times the error on the same data.
The per-body restore fixes it.

Tests

• New: range-spanning prior widths; joined beliefs; sampled materials on the scene base (ranges, values kept across resets, a pinned default restoring each body's own value in a mixed group); joint draws carrying sampled materials into rehearsal and a draw's value reaching only its own world (from-assets play test); the arm refusing to start without the range prior; the menu test.
• Local CI replay of this head: the static checks (yapf, isort, docformatter, mypy, pylint) pass, and 7 of the 8 shards pass.
  Shard 3's one failure (test_preflight_audit.py::test_shadow_tools_pair_outcomes_without_changing_execution, "Not connected to physics server") is the registry flake that Resolve a registry name to its own class, never a subclass that inherits it #213 fixes; master has it too.
  This head with Resolve a registry name to its own class, never a subclass that inherits it #213 applied passes the static checks and all 8 shards.
• CI: static-type-checking fails on every branch until Keep mypy passing on tenacity 9.2's typed retry decorator #212 (tenacity 9.2.1) merges.

🤖 Generated with Claude Code