Keep a declared material's range at least as wide as its plausible range

[yichao-liang]

Stacked on #219 (fit-anchors-never-fitted); review that one first.

Why

Seed 0 of the Domino round fixes_r2, which ran #214's code (merged), lost its test level.
It declared the dominoes' spinning friction on [0, 0.05] and rolling friction on [0, 0.002]; the world has 0.5 and 0.006.
Its level-1 recording, a straight chain, could not identify either, its unfitted prior spanned only the declared boxes, and its rehearsal said 16 of 16 joint draws cascade.
A replay of its episode in the real environment stalls at the world's values and cascades with either friction anywhere inside the declared ranges.
#214 samples only the materials an agent leaves undeclared, so a narrow declaration hides a material just as well.

What changes

• The from-assets arm binds SceneBase with plausible_floor.
  Its parameter registry then names each declared material's plausible range (PLAUSIBLE_MATERIALS, renamed from SAMPLED_MATERIALS; a mass relative to the declared value).
  The range comes from the name and the declaration alone, because the approach stamps the specs before its world binds any observed body.
• stamp_physical_spec_scales widens a declared box narrower than a registry entry's plausible range, so the fit, the prior, the joint draws and the physics sweep all use the wider range.
• The model status lists each widened range, and the scene contract states the ranges (generated from the constant).
• Real-to-sim keeps its declared boxes.

Evidence

A scripted run of the from-assets harness on seed 0's round: level 1 is seed 0's own simulator.py and its recorded actions.
On level 2 it makes seed 0's opening 120-step wait, then rehearses three plans from the current state on 16 joint draws and runs the physics sweep, unfitted as seed 0 played it:

Plan	Joint draws, #214	Sweep, #214	Joint draws, this branch	Sweep, this branch
Seed 0's Oct 8 plan (lost in the world)	16/16	8/23	2/16	10/23
Seed 0's Oct 3 plan (won; bridges 5 mm and 1 degree from the Oct 8 ones)	16/16	8/23	2/16	17/23
The oracle-dynamics arm's plan (won; bridges further out)	16/16	23/23	9/16	21/23

The joint draws now separate the robust plan from seed 0's knife-edge turn, which won once and lost once in the world.
Seven of the Oct 8 plan's 14 failing draws stall the chain; the other seven stall the push, mostly at lateral friction 0.1 to 0.15, which an unfitted prior over the widened box reaches.
The level-1 cascade constrains lateral friction strongly, so a fit (required on test levels by the next PR) narrows that part of the belief.

Tests

• New: the stamping widens a narrower box and keeps a wider box, an open end and a parameter without a range; the scene base names plausible ranges only under the floor, for declared materials only, the same before and after its world binds bodies; the from-assets play test declares a narrow spinning friction and checks the widened spec, the status line, the prior's draws and the sweep, while real-to-sim keeps the declared box.
• 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, for the registry flake it fixes).

🤖 Generated with Claude Code