docs: bench-test runbook — ordered session plan for the bench-pending pile (#122)

[BrettKinny]

Summary

• Adds docs/bench-runbook.md: a single ordered, interactive bench-session plan that merges all eight bench-pending checklists (AXP2101 battery I2C crash — defensive patch shipped #37 State machine — Phase 4 foundation (bench checks) #38 State machine — Phase 5 sleep behaviour (bench checks) #39 State machine — Phase 6 security behaviour (bench checks) #40 Idle empty-room backoff (bench checks) #42 Face-tracking naturalness Phase 0-3 (bench checks) #43 LED redesign round 4 — full right-ring ownership + face indicator (bench checks) #44 LED redesign — round 2/3 (bench checks) #45) from epic Bench-pending epic — items awaiting live device validation #122 into 12 phases (~45–60 min full pass), with per-check columns for what the operator does, what the camera should see, and what the log monitor should see in serial/container logs.
• Translates stale pre-Pi-runtime voice path v1 + full RPi decommission #36 ZeroClaw-era references to the current container architecture (docker restart dotty-bridge, security NDJSON at /var/lib/dotty-behaviour/logs/ inside the dotty-behaviour container, security_cycle.py consumer).
• Flags known risks inline: Phase L overlaps open bug Bridge ↔ firmware state divergence after bridge restart #21 (treated as repro-gathering, not a new failure); the LED redesign round 4 — full right-ring ownership + face indicator (bench checks) #44 dashboard LED-mirror card may be gone post-cutover (N/A-stale, not failed).
• Includes the failure protocol (note, capture logs, keep moving) and the post-session loop: triage → fix via deploy/flash skills → re-run failed IDs only → close out Bench-pending epic — items awaiting live device validation #122.
• Links the runbook from the docs/README.md index.

Test plan

• make doctor references and commands in Phase 0 match the current Makefile targets
• Docs link-check (CI lychee job) passes on the new file
• Dry-read by Brett before the first bench session; correctness of the checklist content is validated by running the session itself

AI assistance: drafted by an AI agent (Claude) from the canonical issue checklists; reviewed by a human before merge.

🤖 Generated with Claude Code