feat(emotion): fit whole-mix V/A calibration on full 60-track corpus

#91 batch finished — all 60 famous tracks fetched + whole-mix analyzed.
First honest baseline on the full corpus (was n=5 before):

  top-1 emotion hit : 22%   (47% w/ overlap)
  V/A quadrant hit  : 42%
  mean V/A error    : 0.573

The center-compression hypothesis is confirmed at scale: 'playful' is the
predicted top emotion for ~18 of 60 tracks (Fleetwood, Marvin, Coltrane,
Avicii, Journey, Eminem...) because the distribution-weighted mean of the
circumplex anchors collapses toward the warm-positive centre.

Affine de-compression (gt ≈ a·pred + b, leave-one-out CV) learns exactly
that bias and corrects it:
  valence:  ×1.03  −0.25   (reads +0.25 too WARM — the playful pull)
  arousal:  ×1.70  −0.39   (compressed ~1.7x — needs the most stretch)
  LOO V/A err 0.573 → 0.519,  quad 42% → 50%.  Modest but real & cross-validated.

calibrate.py now writes per-mode files (calibration.whole-mix.json /
calibration.stem.json) so the #92 whole-mix-vs-stem decision can compare both
without one clobbering the other; calibration.json mirrors the freshest fit.

Stem-aware batch (demucs, all 60) launched in background to settle #92 on the
full corpus, not the n=5 that PLN rightly refused to conclude on.

armada/tide-table/calibration.json

+{
+ "source": "whole-mix",
+ "n": 60,
+ "valence": {
+  "a": 1.0336009052315478,
+  "b": -0.25087256856138784
+ },
+ "arousal": {
+  "a": 1.69672304851225,
+  "b": -0.39451139501465854
+ },
+ "raw": {
+  "va_err": 0.5731779158725924,
+  "quad": 0.4166666666666667,
+  "emo_hit": 0.21666666666666667
+ },
+ "loo": {
+  "va_err": 0.519061292820291,
+  "quad": 0.5,
+  "emo_hit": 0.21666666666666667
+ }
+}
\ No newline at end of file

armada/tide-table/calibration.whole-mix.json

+{
+ "source": "whole-mix",
+ "n": 60,
+ "valence": {
+  "a": 1.0336009052315478,
+  "b": -0.25087256856138784
+ },
+ "arousal": {
+  "a": 1.69672304851225,
+  "b": -0.39451139501465854
+ },
+ "raw": {
+  "va_err": 0.5731779158725924,
+  "quad": 0.4166666666666667,
+  "emo_hit": 0.21666666666666667
+ },
+ "loo": {
+  "va_err": 0.519061292820291,
+  "quad": 0.5,
+  "emo_hit": 0.21666666666666667
+ }
+}
\ No newline at end of file

armada/tide-table/emotion_calibrate.py

@@ -28,9 +28,15 @@ import emotion_ontology as EMO
 HERE = Path(__file__).resolve().parent
 GT = HERE / "emotion_corpus_gt.json"
 ANALYSIS = HERE / "_corpus_cache" / "analysis"
+# Per-mode files so whole-mix and stem calibrations coexist for the #92 decision;
+# `calibration.json` is the symlink/copy of whichever mode we ultimately trust.
 OUT = HERE / "calibration.json"
 
 
+def _out(stem):
+    return HERE / f"calibration.{'stem' if stem else 'whole-mix'}.json"
+
+
 def _pairs(stem=False):
     """(pred_v, pred_a, gt_v, gt_a, slug) for every track that's both GT and analyzed."""
     gt = {t["slug"]: t for t in json.loads(GT.read_text())["tracks"]}
@@ -101,13 +107,16 @@ def fit(stem=False):
     verdict = "✓ helps" if cal_loo[0] < raw[0] else "✗ no gain — need more/cleaner data"
     print(f"\n  {verdict}  (LOO V/A err {raw[0]:.3f} → {cal_loo[0]:.3f})")
 
-    OUT.write_text(json.dumps({
+    payload = json.dumps({
         "source": "stem" if stem else "whole-mix", "n": len(rows),
         "valence": {"a": va_a, "b": va_b}, "arousal": {"a": ar_a, "b": ar_b},
         "raw": {"va_err": raw[0], "quad": raw[1], "emo_hit": raw[2]},
         "loo": {"va_err": cal_loo[0], "quad": cal_loo[1], "emo_hit": cal_loo[2]},
-    }, indent=1))
-    print(f"  ✓ {OUT.name}")
+    }, indent=1)
+    out = _out(stem)
+    out.write_text(payload)
+    OUT.write_text(payload)  # also publish the freshest fit as the active calibration
+    print(f"  ✓ {out.name} (+ {OUT.name})")
 
 
 def apply(v, a):