Research a caster design tolerant of uneven floors

[MJohnson459]

Caster research doc written: design/research/caster.md (follows the lidar_odometry.md style), committed on branch worktree-caster-research (single commit 6da0f15; branch not pushed — no remote git ops without approval).

What it contains:

• Problem framing: the failure is the statically over-constrained 4-contact stance (2 centred wheels + fore/aft casters), not the caster shape — rigid-contact options can't fix it, matching the negative test results with printed/bought ball casters. 3-point stance ruled out (CoM straddles the axle, moves ahead of it with the SO-101 arm).
• Geometry/budget table from repo sources: Ø65 wheels, ~16 mm ride height, casters at ±100 mm, lidar plane at 54 mm; caster envelope 16 mm as-built / ~21 mm with +5 mm wheels / ~27-28 mm with through-plate mounting.
• Survey with verdicts: optimised rigid foot (right tip geometry, insufficient), ball casters/ball transfer units (rejected on evidence), swivel wheel casters (don't fit even the stretched budget; flutter + trail-flip yaw disturbances), sprung drive wheels a la robot vacuums (rejected: major retrofit, chassis still pitches on rigid casters), articulated fore/aft rocker (cleanest theory, no packaging route past the power bank), and the recommendation.
• Recommendation: preloaded plunger casters front AND rear - printed casing (ORP grid) + plunger with shallow-ramp tip, compression spring in a hollow core reacting against the base plate, ~4-5 mm travel, 3-4 N preload, k0.5-1 N/mm. Fits the existing 16 mm ride height: no wheel change, no plate cuts, lidar height unchanged.
• Wobble-vs-compliance analysis with numbers: preload sizing vs CoM offset (~0.3 N.m) and accel load shift (~0.4 N << preload, so contact never breaks); pitch mode ~5 Hz (above <2 Hz drive content); bounded smooth tilt vs today's hard rocking impacts; why 1-2 deg of scan-plane tilt grounds the 54 mm beam at 3.1/1.5 m; why springs beat magnets (magnet repulsion is weakest at ride height, exactly where preload is needed - magnets only OK pre-compressed).
• BOM delta: ~GBP 6-10 (spring assortment ~GBP 6, alt Ø12x3 N42 magnet pairs ~GBP 5, optional PTFE tip inserts ~GBP 4) and the implied CAD/mounting changes for the follow-up task (retire Caster.step, two-part casing+plunger x2, ASSEMBLY/BOM/URDF updates, rug-edge acceptance test).

Verification: no CAD/STL/3MF touched (deliverable is the doc only, as the issue requires); geometry cross-checked against robot.yaml, the URDF, and Caster.step (bounding box measured from the STEP file); design/ is pre-commit-excluded so lint is clean by construction.