j133 strafe fire detection: correct the enemy energy delta for the server's +3*power hit bonus and our own damage, never drop a too-large drop (TR_STRAFE_FIRE_FIX, default on); catch 98.89%->100% of enemy fires on the 70-battle corpus

common_libs/tests/label_inversion_three_way.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Task C — the label-inversion question under ONE consistent computation (j133).
+
+Reproduces, on the recorded corpus and with the SAME extraction/metric, the
+correlation between each candidate DANGER MAP and the realised per-bin hit rate
+`P(hit | b_our = g)` (the j128 metric, whose histogram value is -0.342):
+
+  (i)   histogram label         danger(g) = P(arrival bin = g)             [j128]
+  (ii)  outcome proxy label     danger(g) = P(hit and |g - b_our| <= w)    [j130]
+  (iii) EXACT bullet-line label danger(g) = P(|g - b_bullet| <= w)         [j131]
+  (iv)  the STATE-CONDITIONAL outcome model's own predicted danger, held out
+        by battle  danger(g) = mean_test P_hat(hit | state, g)             [new]
+
+Negative = minimising the danger steers INTO where the observed hits happen.
+If the physically-EXACT label (iii) is still negative, exact geometry does NOT
+fix the inversion and the observable STATE is the binding constraint.
+
+Run:
+  python3 common_libs/tests/label_inversion_three_way.py --corpus /tmp/tfil_ab2/out
+"""
+from __future__ import annotations
+
+import argparse
+import os
+import statistics
+import sys
+
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+
+import analyze_drussgt_dodge_vs_power as adp
+import outcome_label_gate as olg
+
+NBINS = olg.NBINS
+
+
+def realised(recs):
+    n = [0] * NBINS
+    h = [0.0] * NBINS
+    for r in recs:
+        n[r["b_our"]] += 1
+        h[r["b_our"]] += r["hit"]
+    used = [b for b in range(NBINS) if n[b] > 0]
+    rate = [h[b] / n[b] for b in used]
+    return used, rate
+
+
+def label_corrs(recs):
+    used, rate = realised(recs)
+    hist = statistics.correlation([sum(1 for r in recs if r["b_our"] == b) / len(recs)
+                                   for b in used], rate)
+    proxy = statistics.correlation(
+        [statistics.fmean(olg.hitwin(r, b) for r in recs) for b in used], rate)
+    exact = statistics.correlation(
+        [statistics.fmean(1 if abs(b - r["b_bullet"]) <= r["w"] else 0
+                          for r in recs) for b in used], rate)
+    return hist, proxy, exact
+
+
+def model_corr(recs, seed, state_free):
+    """Held-out (split BY BATTLE) state-conditional model danger vs test hit rate."""
+    tr_b, te_b = olg.split_battles({r["battle"] for r in recs}, seed)
+    tr = [r for r in recs if r["battle"] in tr_b]
+    te = [r for r in recs if r["battle"] in te_b]
+    om = olg.OutcomeModel(decay=128, shift=1, state_free=state_free)
+    edges = dict(olg.CANON)
+    for r in tr:
+        st = olg.code_of(r, edges)
+        for g in range(NBINS):
+            om.learn(st, g, olg.hitwin(r, g))
+    used, rate = realised(te)
+    danger = [statistics.fmean(om.predict_hit(olg.code_of(r, edges), b) for r in te)
+              for b in used]
+    return statistics.correlation(danger, rate)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--corpus", default="/tmp/tfil_ab2/out")
+    ap.add_argument("--seeds", type=int, default=3)
+    args = ap.parse_args()
+
+    recs = olg.extract(adp.discover_tfil(args.corpus))
+    hist, proxy, exact = label_corrs(recs)
+    ms = [model_corr(recs, s, False) for s in range(args.seeds)]
+    mf = [model_corr(recs, s, True) for s in range(args.seeds)]
+
+    print(f"corpus: {args.corpus}  records: {len(recs)}  "
+          f"base hit: {statistics.fmean(r['hit'] for r in recs) * 100:.2f}%")
+    print()
+    print("corr( danger(g) , P(hit | b_our = g) )  [the j128 metric]")
+    print("------------------------------------------------------------")
+    print(f"(i)   histogram label (j128)          : {hist:+.3f}")
+    print(f"(ii)  outcome proxy label (j130)      : {proxy:+.3f}")
+    print(f"(iii) EXACT bullet-line label (j131)  : {exact:+.3f}")
+    print(f"(iv)  state-CONDITIONAL outcome model : {statistics.fmean(ms):+.3f} "
+          f"(seeds {['%+.3f' % v for v in ms]})")
+    print(f"      state-FREE outcome model        : {statistics.fmean(mf):+.3f} "
+          f"(seeds {['%+.3f' % v for v in mf]})")
+    print()
+    if exact < 0:
+        print("VERDICT: the physically-exact label is STILL negative -> the LABEL "
+              "was never the problem; the observable STATE is the binding "
+              "constraint (closes the learned-movement family).")
+    else:
+        print("VERDICT: the exact label is positive -> geometry, not state, was "
+              "the binding constraint.")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())