shakedown: wave 10: MSA repeatability harness (gage error vs tolerance)

[ugglr]

The Measurement System Analysis leg of the production-QA path: before trusting a pass/fail, characterize the measurement itself.

What this adds

• Verification/scripts/repeatability.sh. Runs the variance phase (Phase 4) N times back to back on one unit and reports the gage's own run-to-run noise as a fraction of the fail tolerance: 6-sigma of spread_pct over the runs, divided by the 10-point fail band, read against the AIAG Gage R&R rule (under 10% capable, 10-30% marginal, over 30% the gage rivals the limit it decides against). Repeatability only (one unit, back to back); reproducibility across stations/days is a separate, later step. REPEAT_N (default 5); every cpu-variance.sh knob passes through.
• Docs: Production QA.md MSA row moves Absent -> Partial and leg 2 from "(next)" to "(this wave)"; README layout + roadmap line; CHANGELOG.

Why

A "10% spread" fail limit is only meaningful if our own measurement noise is well inside it. This harness answers "is the metric even capable of a pass/fail." It already pays off: at very short iteration settings the gage comes back inadequate, which is the honest finding that a one-second iteration is not a measurement.

Verification

Lint clean (shellcheck, heredoc ast.parse, JSON valid, zero em dashes). Ran end-to-end on this Intel box (3 fast runs): mechanism works, JSON contract matches the doc, the capability buckets and the < 2 valid runs / division / trap-cleanup paths all behave. Passed dad review (correctness clean, consistency one-voice, simplicity flagged nothing cut-before-ship).

Still owed for production-grade MSA (documented, not in this wave): the actual characterization run on a golden unit, and reproducibility across stations and days.