P2 validated on 2141 vendor GTCs (1000 Genomes Omni2.5)

Raw X/Y and controls assembled from IDATs equal the vendor GTC arrays for
2138/2141 samples x 2.45M loci; 3 single-locus differences are recorded
as K-009 (probable per-channel missing rule in AutoCall 1.6.2.2). Channel
order X=Red confirmed (U1). `simurg ingest` processed all 2141 samples in
19 min (I/O-bound). AWS run cost ~$0.34 and left no resources.

docs/validation/P2.md

@@ -0,0 +1,39 @@
+# Phase P2 validation report — canonical model and raw X/Y assembly
+
+Date: 2026-09-27. Local work on the Pi; large run on a self-destructing AWS spot instance
+(`run-20260927T195800Z`, m7g.4xlarge, 77 min, ≈ $0.34). Summary: `P2_omni2141_summary.json`.
+
+## Delivered
+
+| Item | Module |
+|---|---|
+| Raw X/Y + control assembly (plan §7.1), per-chip address-map cache | `signal/raw_xy.py` |
+| IDAT ↔ manifest address-coverage gate (IN-027) | `signal/raw_xy.py` |
+| Canonical Arrow schemas, deterministic Parquet (plan §5) | `core/schemas.py` |
+| `simurg ingest` (discovery, sheet, naming, overwrite/resume, workers, run manifest) | `pipeline/ingest.py`, `cli/main.py` |
+
+## Acceptance (plan §18 P2: raw X/Y exactly equal to vendor GTC TOC 1000/1001, 100 % loci × ≥ 50 samples)
+
+| Check | Data | Result |
+|---|---|---|
+| Raw X/Y and controls vs oracle GTCs | 18 GSA v4 samples (gtc2vcf GTCs) | **EXACT, 18/18** |
+| Raw X/Y vs **vendor** GTCs (AutoCall 1.6.2.2) | 2141 Omni2.5 samples × 2 450 000 loci | **2138/2141 samples EXACT**; 3 loci in 3 samples differ (below) |
+| Controls vs vendor GTCs | 2141 samples | **EXACT** |
+| Channel assignment (plan U1): X = Red, Y = Grn for Infinium II | 5.2 × 10⁹ locus values | **confirmed** (0 swapped loci) |
+| `ingest` outputs identical across thread counts and layouts | synthetic + GSA | **byte-identical** |
+
+The gate "100 % loci × ≥ 50 samples" is met (2138 samples fully exact). The 3 remaining differences are one locus
+each in samples 5767788121_R01C01, 5826354152_R02C01 and 6268201099_R02C01. In each, we mark the locus missing
+(0, 0) but the vendor kept **one** channel (e.g. `rs178512`: vendor `[7086, 0]`). Hypothesis K-009: this AutoCall
+version zeroes each channel independently when that channel's bead count is < 2, while the modern rule (gtc2vcf,
+emulating current vendor software) zeroes both. Not yet confirmed (needs the bead counts of those three IDATs).
+
+## `simurg ingest` at scale (2141 Omni2.5 samples, 16 workers)
+
+| Metric | Value |
+|---|---|
+| Wall time | 18 min 52 s (0.53 s/sample) |
+| CPU utilisation | 397 % of 1600 % — I/O-bound: ~120 GB of IDATs read through a gp3 volume at its 125 MB/s baseline |
+| Peak RSS (main process) | 2.4 GB |
+| Output | 19 GB (≈ 9 MB/sample) |
+| Status | 2141 OK; 285 × `W_IDAT_LOW_ADDRESS_COVERAGE` (0.03 % of addresses absent), 4 × `W_IDAT_NO_SCAN_TIME` |