[HyorinmaruK]

…ix -> assembly

• The whole first batch is defective (wrong tool offset). The camera counts defects; N among the last 2N inspected parts (website default 6) stop the whole line, both machines at once (halt_scope='line').
• Quarantine (contain_suspect): every part made with the bad program goes to the caliper chamber, flagged or not, including parts already in the buffer.
• The chamber writes chamber_measurements.csv in the physical caliper template format (tier SIMULATED, never accepted as physical evidence); the CNC agent reads it to compute the G-code correction; operator approval kept.
• No blank is cut during a stop (a machine reacquired after a stop waits).
• Machined coupons: crack_p=0. Seeds 0-2: 0 escapes, 60/60 shipped good.
• Website: line-stop banner, defects counted N/N, quarantined count, CSV rows with download, CNC agent G-code diff, new tour, N input defaults to 6.
• The per-machine variant stays available (halt_scope='machine').

Claude-Session: https://claude.ai/code/session_011qYWELVQUDWGU4JdLmjN7C

What and why

Evidence

Checklist

• python -m pytest -q passes locally
• ruff check ., mypy qualityrelay17, python scripts/check_schemas.py pass (CI runs them too)
• API changed? OpenAPI snapshot regenerated with python scripts/check_schemas.py --fix
• No keys, tokens, .env or personal photos added
• Claims in docs/README match what was actually measured (no VERIFIED/WORKING without evidence)

Safety-critical? (paths in .github/safety-paths.txt)

• This PR touches actuation, auth/keys/signatures, the ledger/verification, SAFE-MESH, the agent-message bus, firmware, or CI/review rules
  • Needs two approvals from people other than the author, on the latest commit (safety-review-gate check). Pushing again resets it.
  • Describe the failure mode you considered and the test that pins it:

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
qualityrelay	Ready	Preview	Nov 1, 2026 5:05pm UTC