fix(site): subset-ness is relative to the base, and each scoped run is screened on its own flags

1. isStrictSubset compared against `widest` rather than the run being layered
   ONTO. Subset-ness only means anything relative to the base: a scoped re-run
   is a subset of the GRID, not necessarily of some historical snapshot. Add a
   7th dialect, re-run OTB_DIALECTS=openai,vertex, and its 4 cells contain
   vertex|openai - a coord the old 36-cell run never had. Not a subset of
   widest, so it landed in `fulls`, became the base by recency, and 33 measured
   cells vanished. That is the exact failure this module exists to prevent,
   arriving through the door next to the one it was watching.

2. The degraded-mode refusal read snap.matrix, which by then is the LAYERED
   result carrying the BASE run's mode flags - layerScopedMatrix copies cells
   and nothing else. Every scoped snapshot that layered in was exempt, so a
   verify-local smoke run with cell_perf_sweep: false could replace the headline
   cell of a clean field run while the guard read the base's `true` and threw
   nothing. Screened per run now, which is the only place the scoped run's own
   flags still exist.

The 3 red site tests are C8 only - the board mixes two engines while the field
run is mid-flight. They clear when it completes.

Author: MattJackson

site/gen-data.mjs

@@ -230,12 +230,36 @@ function resolvedSnapshot(key) {
   // The base is the newest run that is NOT a strict subset of the widest - i.e. a full run.
   const fulls = snaps.filter((s) => !isStrictSubset(s.__coords, widest.__coords));
   const base = fulls[fulls.length - 1];
-  const scopedNewer = snaps.filter((s) => isStrictSubset(s.__coords, widest.__coords) && s.__ms >= base.__ms);
+  /* SUBSET OF THE BASE, NOT OF THE WIDEST. Subset-ness only means anything relative to the run being
+     layered ONTO: a scoped re-run is a subset of the GRID, not necessarily of some historical
+     snapshot. Testing against `widest` misclassified any re-run that touched a coordinate the widest
+     run lacked - add a 7th dialect, re-run `OTB_DIALECTS=openai,vertex`, and its 4 cells contain
+     `vertex|openai`, which the old 36-cell run never had. Not a subset of `widest`, so it landed in
+     `fulls`, became the base by recency, and 33 measured cells vanished from the board. That is the
+     exact failure this module was written to prevent, arriving through the door next to the one it
+     was watching. */
+  const scopedNewer = snaps.filter((s) => isStrictSubset(s.__coords, base.__coords) && s.__ms >= base.__ms);
   if (!scopedNewer.length) return base;
   const out = structuredClone(base);
   out.__file = base.__file;
   let m = base.matrix;
-  for (const sc of scopedNewer) m = layerScopedMatrix(m, sc.matrix, sc);
+  for (const sc of scopedNewer) {
+    /* AND EACH SCOPED RUN IS SCREENED ON ITS OWN FLAGS. The degraded-mode refusal below reads
+       `snap.matrix`, which by then is the LAYERED result carrying the BASE run's mode flags -
+       `layerScopedMatrix` copies cells and nothing else. So every scoped snapshot that layered in was
+       exempt from the guard, and a `verify-local` smoke run with `cell_perf_sweep: false` could
+       replace the headline cell of a clean field run while the guard read the base's `true` and threw
+       nothing. Screening here, per run, is the only place the scoped run's own flags still exist. */
+    const degraded = snapshotDegradedMode(sc.matrix);
+    if (degraded) {
+      throw new Error(
+        `gen-data: REFUSING to layer a DEGRADED-MODE re-run into ${key}. ${sc.__file} ` +
+        `(measured_at ${sc.measured_at}) ran with ${degraded} - those phases were switched OFF, so it ` +
+        `is a local smoke run, not a measurement, and its cells would replace measured ones.\n` +
+        `  Fix: delete or move that snapshot out of results/snapshots/.`);
+    }
+    m = layerScopedMatrix(m, sc.matrix, sc);
+  }
   out.matrix = m;
   out.__layered_from = scopedNewer.map((s) => s.__file);
   return out;