SmolVLA for Bob: the first learned manipulation skill, through the gap rail

## Why
The 2026 open-weights VLA field sorts cleanly for Bob's hardware: SmolVLA (~450M, LeRobot-native, benchmarked on physical SO-101, 60–75% of OpenVLA at 1/14th the size, closing to 5–10% after single-task fine-tuning with ~200 demos; ≥10 Hz on consumer hardware). It is a POLICY under LeRobot, not an Ollama pull — a gap-rail skill, the biggest single project on this list. Expect this issue to spawn sub-issues.

## What (phased — the marathon scopes phase 1, maybe 2)
1. **Record**: teleop demo capture through the app's existing arm controls → LeRobot dataset format on the Pi (every commanded action already flows through the signed gateway; recording is a tap)
2. **Train**: fine-tune SmolVLA off-Pi (rented GPU-hour); document the loop
3. **Serve**: smolvla actuator-side runtime on the Pi emitting joint targets INTO the gateway path — every action judged, tier-gated, receipted, deadman-covered like any other; declared via `castor capability add` (operator-signed)

## Acceptance (phase 1)
- 10 recorded demos of one pick-place task replayable from the dataset; format loads in LeRobot
## Authority rule
The policy drafts motion; the gateway remains the exclusive path to the servo bus. A VLA that could bypass the gateway is a non-starter — same rule as everything else in this stack.

## Pointers
- castor/hardware/so_arm101/* (safety_bridge, rcan_bridge), docs/SKILL-GAPS.md (the rail), Bob gripper manifest-zero lesson (calibration_api docstring)

Read castor/hardware/so_arm101/*, especially safety_bridge and rcan_bridge, along with docs/SKILL-GAPS.md and the calibration_api docstring lesson. Start by tracing the app's existing arm controls and how commanded actions reach the gateway, then determine how phase 1 recording maps to LeRobot format. Done means 10 demos of one pick-place task are replayable from a dataset that loads in LeRobot.

Written by the indexing model from the issue text.