Add quality automation spec: stamper metrology and layered record QC

- Keyence-class instrument stack for father/mother/stamper analysis
  (VK-X confocal sampled zones, VR-6000 wide-area, XRF thickness) with
  genealogy integration via middleware
- Three-layer pressed-record QC: 100% inline optical inspection,
  per-cycle press telemetry, sampled automated audio comparison
- Honest limits documented: no turnkey products exist for the optical
  and audio layers; 100% audio QC impossible at press speed

docs/quality-automation.md

@@ -0,0 +1,130 @@
+# Quality Automation
+
+The target is to automate as much as possible. For quality control this
+means instruments and inspection stations that measure, decide, and log into
+the [traceability system](traceability.md) without manual transcription.
+Research findings from vendor sites and trade sources, July 2026; the
+recommended architectures are engineering positions to be validated with
+vendors, not purchased designs.
+
+## Stamper, father, and mother metrology
+
+What must be measured on the electroformed nickel parts (~0.2 to 0.3 mm
+thick): groove geometry (grooves are 40 to 80 microns wide, ~28 microns
+deep, 200 to 350 grooves per inch, inverted to ridges on a stamper),
+thickness uniformity (total thickness variation of 10 to 20 microns is the
+achievable target, by analogy with CD stamper practice), surface roughness
+on the groove walls, and micro-defects.
+
+A Keyence-class instrument stack covers this, in tiers:
+
+- **Keyence VK-X 3D laser confocal microscope**: 20 nm vertical and ~120 nm
+  lateral resolution, non-contact. The reference-grade instrument for groove
+  cross-sections, wall roughness, and fine defect analysis. Scanning the
+  full spiral (hundreds of meters of groove or more) is not feasible; the
+  correct strategy is **sampled radial zones** (lead-in, outer, mid, inner
+  radius). The optical disc industry verified its nickel stampers with the
+  same instrument class, so the technique transfers directly.
+- **Keyence VR-6000 wide-area profiler**: 0.1 micron height resolution,
+  ~1 second per field, with an HDR mode built for glossy surfaces (bright
+  nickel is near-mirror, the classic hard case for optics; manageable, but
+  plan tuning time). Fast wide-area form, flatness, and gross defect
+  mapping.
+- **XRF thickness gauge** (Fischer/Bowman class): non-destructive nickel
+  deposit thickness per serial, replacing the industry's manual 10-point
+  micrometer practice. Thickness must be read on the outer land/shoulder
+  zones, not across the groove field.
+- **Stylus profilometer** kept as the traceable calibration reference.
+- **Inline machine vision** for fast 100% cosmetic screening of every part.
+
+Integration is real but not turnkey: the Keyence instruments export CSV and
+raw 3D data and support scripting/REST layers, and the inline sensors speak
+Ethernet/IP, but there is no off-the-shelf MES connector. Budget a light
+middleware service that stamps each measurement with the part serial and
+batch ID and writes to the genealogy database.
+
+Industry context: even a respected stamper house checks thickness with a
+hand micrometer at 10 points on sacrificial blanks, with nothing logged.
+Automated non-contact metrology with per-serial logging is ahead of current
+industry practice, which is exactly the point: it makes every father, mother,
+and stamper a data-backed accept/reject decision before the next plating
+generation or press mounting.
+
+## Pressed record QC: three layers
+
+Full automated audio checking of every record is physically impossible: a
+12" record plays for ~40 minutes both sides while a press makes a new record
+every ~25 seconds, so one press outruns roughly 80 to 95 full-time listening
+stations. No commercial playback-QC robot exists anywhere. The credible
+maximum-automation architecture is layered:
+
+### Layer 1: 100% optical inspection, inline
+
+Machine vision can inspect every record at line speed. Detectable with
+darkfield lighting and line-scan cameras: non-fill (dull patches where PVC
+did not fill the stamper), stitching and flow lines, warp (via structured
+light or profilometry), off-center spindle holes (concentricity of hole vs
+groove spiral), surface debris, and label errors (wrong label, off-center,
+print defects). Non-fill and stitching correlate with audible defects, so
+optical inspection is a strong proxy for audio quality without playing the
+record.
+
+No vinyl-specific vision product exists; this is a custom integration of
+general machine vision (Keyence or Cognex class) at the press unloader. CNX
+would be an early adopter, and it should be budgeted as an integration
+project.
+
+### Layer 2: press telemetry, every cycle
+
+The proven, deployed piece. Viryl's ADAPT platform demonstrates the model:
+per-cycle nozzle pressure, mold temperatures, cycle time, cake weight,
+yields, and drift alerts streamed from the press. The Pheenix AD12 runs a
+Beckhoff PLC exposing similar data over standard industrial protocols, and
+supports remote monitoring. Machine data access is already flagged as a
+procurement requirement in the traceability spec.
+
+### Layer 3: automated audio comparison, sampled
+
+A needle-drop capture station with automated analysis, run on samples:
+every test pressing / first-off per stamper, plus pulled samples scaled by
+stamper age (stampers last roughly 1,000 to 2,000 records; sampling should
+densify near end of life; industry rule of thumb is a spot-play around every
+50th record).
+
+The analysis pipeline is buildable from proven parts but is a first-of-kind
+system as a whole:
+
+1. Capture the needle-drop (a conventional tonearm; laser pickups exist but
+   are dust-sensitive and impractical as line instruments)
+2. Spectral-envelope comparison against the source master (the approach
+   proven by Stamper Discs' Fonograph tool, which avoids the hard
+   time-alignment problem by comparing whole-program FFT fingerprints)
+3. Click/pop detection and counting (mature algorithms of the iZotope RX
+   class, scriptable)
+4. Noise floor from lead-in/run-out silence, channel balance from program
+   RMS
+5. Numeric scores and pass/fail to the traceability database
+
+What only this layer can catch: wrong master cut into good-looking grooves,
+EQ/mastering errors, and speed/pitch problems. Optical and audio layers are
+complements, not substitutes.
+
+## The traceability payoff
+
+The natural record identity is **stamper ID + press ID + cycle timestamp +
+side**. Against that key: press telemetry on every cycle, an optical result
+on every record, an audio score on every sampled record, and the stamper's
+own metrology history from before it was mounted. Audio scores trending down
+against a stamper's play count become an automatic retirement signal, and a
+customer complaint about one record resolves to a specific stamper, press,
+shift, and compound lot.
+
+## Honest limits
+
+- No vendor sells a turnkey version of layers 1 or 3, or the metrology
+  middleware; all three are integration/development projects
+- 100% audio QC of every record is not achievable at press speed and should
+  never be promised to customers; 100% optical plus sampled audio is the
+  accurate claim
+- Human listening stays in the loop for test-pressing approval; the
+  automation makes sampling systematic and logged, it does not replace ears

docs/traceability.md

@@ -21,7 +21,8 @@ For any order, the customer can follow the full chain:
    parameters (temperatures, pressures, cycle times), PVC compound lot and
    virgin/recycled mix, label lot
 6. **Quality control**: test pressing approvals, inspection results, reject
-   counts and reasons
+   counts and reasons (the automated instrument layers feeding this are
+   specified in [quality automation](quality-automation.md))
 7. **Finishing and packaging**: trimming, sleeving, carton lot, packaging
    partner batch
 8. **Fulfillment**: warehouse location, shipment or dropshipping consignment,