The validator was a hand re-implementation that had drifted from
scorer.ex/band_config.ex/region.ex, so every rho in
docs/algo-reports/validation-*.md measured an algorithm that does not
exist:

- pressure was inverted (88 at low MSLP falling to 30 at high, vs
  production's ascending 30 -> 88), including the delta curve
- score_season omitted the Region multiplier
- time-of-day used a month-only sunrise table and the old fixed 8 pm
  evening cutoff, not BandConfig's latitude/month sunrise + sunset
- rain/humidity/seasonal/refractivity constants were pre-2026-04 values
- weights were not resolved through BandWeights' JSON-first lookup

Rebuilt from the source of truth and re-derived every band table; the
whole port is now value-for-value identical to production (verified
against a dump of the real scorer over every band, factor grid and the
Region boxes).

The baseline table was also degenerate: pers_deltas was literally 0.0,
so rho(persistence) == 0 and 'skill gain' == rho(algorithm), while
climatology and no-skill shared a rank vector. Predictors are now each
scored against the raw distance target, so 'does the score beat
month-of-year alone?' is a direct rho comparison; a constant predictor
reports an undefined rho (null / em dash), never a fabricated 0.0, and
the seasonality-removed view is kept for the algorithm only.

Three factors (sky/wind/rain) have no HRRR column in the corpus, so they
stay pinned to their neutral constants; the report and JSON now state
that share of composite weight explicitly instead of hiding it.

Also aligns the flagged_invalid predicate with the column definition
(boolean NOT NULL DEFAULT false), ports the retired duct/Richardson
inputs so their inertness is proven rather than assumed, and refreshes
the module docstring.

@@ -152,2 +275,4 @@
 
 def score_humidity(abs_hum: float | None, band_cfg: dict) -> int:
+    # Divergence, deliberate: production has no nil clause here. A nil
+    # abs_humidity falls through Erlang term ordering (`nil < 4` is true) to 55

@@ -441,0 +818,4 @@
+    return {
+        "n": n,
+        "alg_rho": _rho4(alg_rho),
+        "pers_rho": _rho4(pers_rho),

@@ -151,6 +274,11 @@ def c_to_f(c: float) -> float:
 
 
 def score_humidity(abs_hum: float | None, band_cfg: dict) -> int:
+    # Divergence, deliberate: production has no nil clause here. A nil

@@ -177,0 +350,4 @@
+        return 78  # "Good — inversion eroding"
+    if -3.0 <= delta < -1.5:
+        return 82  # "Pre-dawn — inversion building"
+    if local >= sunset and local - sunset <= 3.0:

@@ -441,0 +822,4 @@
+        "clim_rho": _rho4(clim_rho),
+        "no_skill_rho": _rho4(ns_rho),
+        "alg_rho_anomaly": _rho4(alg_rho_anomaly),
+        "skill_gain": _rho4(skill_gain),

@@ -0,0 +1 @@
+/Users/graham/dev/ntms/prop/deps

@@ -28,6 +28,7 @@ defmodule Microwaveprop.Propagation.ValidationReportTest do
         report["per_band"]
         |> Enum.filter(fn {_band, v} -> v["n"] >= 50 end)
         |> Enum.map(fn {_band, v} -> v["skill_gain"] end)
+        |> Enum.reject(&is_nil/1)

@@ -671,0 +682,4 @@
+        # An undefined ρ on either side is not evidence of a regression — there
+        # are no two numbers to compare, so the band keeps the candidate's
+        # weights rather than being reverted on a measurement that does not exist.
+        if new_rho is None or old_rho is None: