Retrain heuristic weights (2026-08-21)

[github-actions]

Automated re-fit. Review the weight diff below, then update the golden tests
in HeuristicArenaDataSourceTests with the printed values before merging.

Golden gate (run in the retrain job — a GITHUB_TOKEN PR does not trigger build):
FAIL — goldens must be updated from the trainer output before merging

training dir: D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training
live fetch; snapshotting payloads to D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training\.snapshot
fetching HSReplay arena card stats...
HSReplay pooled (card,class) rows: 2187
rows: 2187 (single source)
features: 32 fitted, 18 dropped
  dropped: kw_charge(n=11), kw_colossal(n=11), kw_combo(n=21), kw_echo(n=6), kw_forge(n=3), kw_freeze(n=23), kw_magnetic(n=14), kw_outcast(n=18), kw_poisonous(n=35), kw_spellpower(n=19), kw_stealth(n=10), kw_windfury(n=8), tx_armor(n=28), tx_destroy_minion(n=29), tx_discover(redundant), tx_restore_amt(n=18), tx_silence(n=17), tx_transform(n=32)
penalty BEFORE normalizing weights: alpha=10 vs mean(diag)=75867  -> shrinkage 0.000132, df_eff=31.0/32
CV: 5-fold grouped by card (1087 cards, 2187 rows), 5 repeats, 11 alphas

     alpha   withinClassRho       se
      0.03           0.2132   0.0035
       0.1           0.2132   0.0035
       0.3           0.2134   0.0035
         1           0.2138   0.0035
         3           0.2145   0.0032
        10           0.2174   0.0026
        30           0.2192   0.0021
       100           0.2207   0.0018
       300           0.2163   0.0020
      1000           0.2014   0.0033
      3000           0.1775   0.0028
CV best alpha=100 (rho=0.2207); 1-SE rule picks alpha=100 (rho=0.2207, threshold 0.2189)
penalty AFTER: alpha=100 vs mean(diag)=2381  -> shrinkage 0.040306, df_eff=28.2/32

holdout evaluation (pooled Spearman on held-out rows, alpha=100):
  random cards (baseline)            n= 108  rho= 0.2863  (se 0.0288)
  lowest-games decile                n= 108  rho= 0.1390  (up to 197 games)
  -> covariate-shift gap vs baseline: -0.1473
  target spread / model error:
      thin  sd(y)=6.904  MAE=5.871
      rest  sd(y)=5.044  MAE=3.272
  calibration slope (truth regressed on prediction):
      thin decile  0.8108   random holdout  0.8427  (se 0.1509)
      -> measured ModelOnlyShrink = 0.9622 (thin/random; set the runtime constant from THIS, not from a correlation ratio)
  leave-out set CORE                 n= 207  rho= 0.1878
  leave-out set TIME_TRAVEL          n= 168  rho= 0.3785
  leave-out set BATTLE_OF_THE_BANDS  n= 158  rho= 0.2990
  leave-out set WHIZBANGS_WORKSHOP   n= 153  rho= 0.2461

fitting ridge (alpha=100) on 2187 rows, 32 features...
display anchor (pool median raw): -1.12, robust sigma: 1.3521 -> 15 pts per robust SD
wrote D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training\arena_weights.generated.json

feature                  committed   generated     diff
intercept                   +1.09      +0.60    -0.49
anchor_median_raw           -0.92      -1.12    -0.20  <-- differs
attack                      -0.15      -0.20    -0.05  <-- differs
cost                        +0.15      +0.00    -0.15  <-- differs
has_tribe                   +0.12      +0.50    +0.38  <-- differs
health                      -0.18      -0.20    -0.02  <-- differs
is_hero                     +0.81      +4.04    +3.23  <-- differs
is_loc                      -0.88      +0.12    +1.00  <-- differs
is_minion                   +2.61      +1.52    -1.09  <-- differs
is_neutral                  -1.94      -2.00    -0.06  <-- differs
is_spell                    -2.75      -1.72    +1.03  <-- differs
is_weapon                   -0.94      -0.24    +0.70  <-- differs
kw_battlecry                +0.70      +0.72    +0.02  <-- differs
kw_deathrattle              -0.41      -0.81    -0.40  <-- differs
kw_discover                 +0.86      +0.73    -0.13  <-- differs
kw_divine_shield            +0.10      -0.05    -0.15  <-- differs
kw_lifesteal                +0.84      +0.69    -0.15  <-- differs
kw_reborn                   +1.83      +1.16    -0.67  <-- differs
kw_rush                     +1.05      +1.02    -0.03  <-- differs
kw_secret                   -0.19      -1.20    -1.01  <-- differs
kw_taunt                    -0.10      +0.06    +0.16  <-- differs
kw_tradeable                +0.71      +1.07    +0.36  <-- differs
stat_per_mana               -2.40      -1.33    +1.07  <-- differs
statline                    +0.34      +0.18    -0.16  <-- differs
tx_aoe                      +0.51      +0.52    +0.01  <-- differs
tx_damage_amt               +0.08      +0.13    +0.05  <-- differs
tx_dmg_per_mana             +0.00      -0.28    -0.28  <-- differs
tx_draw                     -0.31      -0.07    +0.24  <-- differs
tx_gain_card                -0.87      -0.65    +0.22  <-- differs
tx_mana_cheat               +1.05      +0.61    -0.44  <-- differs
tx_persistent               +0.15      +0.48    +0.33  <-- differs
tx_random                   +0.32      +0.23    -0.09  <-- differs
tx_summon                   +0.95      +1.13    +0.18  <-- differs
weapon_value                -0.20      -0.35    -0.15  <-- differs

max abs weight diff: 3.23
weights differing by >= 0.01: 33
GATE: generated weights are close but not identical (see rows above).

golden scores for HeuristicArenaDataSourceTests.Goldens
(paste the VALUES over the literals on adopt; keep the trailing comments):
  { HSCard.PlatedBeetle, 24.63 },
  { HSCard.ElvenArcher, 50.44 },
  { HSCard.DefenderOfArgus, 40.95 },
  { HSCard.FieryWarAxe, 51.44 },
  { HSCard.Fireball, 54.93 },
  { HSCard.ColdlightOracle, 46.89 },
  { HSCard.DireFrenzy, 50.00 },
  { HSCard.NorthshireCleric, 54.10 },
  { HSCard.DarkIronDwarf, 28.39 },
  { HSCard.Deathwing, 0.00 },

bootstrapping 300 card-resampled refits...

pick instability under resampling, by how clearly the model named a winner:
  winner leads by >  0.0 pts:  27.84% of those picks flip
  winner leads by >  1.0 pts:  25.73% of those picks flip
  winner leads by >  3.0 pts:  21.64% of those picks flip
  winner leads by >  5.0 pts:  18.25% of those picks flip
  winner leads by > 10.0 pts:  11.20% of those picks flip

feature                    weight       se    |w|/se   signCons
attack                      -0.20     0.09       2.2       0.98
cost                        +0.04     0.08       0.4       0.68  <-- within noise
has_tribe                   +0.50     0.25       2.0       0.98  <-- within noise
health                      -0.20     0.09       2.3       0.99
is_hero                     +4.04     3.01       1.3       0.79  <-- within noise
is_loc                      +0.12     0.96       0.1       0.57  <-- within noise
is_minion                   +1.52     0.71       2.1       1.00
is_neutral                  -2.00     0.29       7.0       1.00
is_spell                    -1.72     0.67       2.6       1.00
is_weapon                   -0.24     1.10       0.2       0.59  <-- within noise
kw_battlecry                +0.72     0.29       2.5       1.00
kw_deathrattle              -0.81     0.31       2.7       1.00
kw_discover                 +0.73     0.29       2.5       0.99
kw_divine_shield            -0.05     0.56       0.1       0.53  <-- within noise
kw_lifesteal                +0.69     0.45       1.5       0.93  <-- within noise
kw_reborn                   +1.16     1.03       1.1       0.87  <-- within noise
kw_rush                     +1.02     0.36       2.8       1.00
kw_secret                   -1.20     0.56       2.1       0.98
kw_taunt                    +0.06     0.35       0.2       0.55  <-- within noise
kw_tradeable                +1.07     0.54       2.0       0.97  <-- within noise
stat_per_mana               -1.33     1.06       1.3       0.85  <-- within noise
statline                    +0.18     0.07       2.6       0.99
tx_aoe                      +0.52     0.36       1.4       0.94  <-- within noise
tx_damage_amt               +0.13     0.12       1.1       0.87  <-- within noise
tx_dmg_per_mana             -0.28     0.46       0.6       0.68  <-- within noise
tx_draw                     -0.07     0.37       0.2       0.52  <-- within noise
tx_gain_card                -0.65     0.34       1.9       0.97  <-- within noise
tx_mana_cheat               +0.61     0.40       1.5       0.93  <-- within noise
tx_persistent               +0.48     0.32       1.5       0.92  <-- within noise
tx_random                   +0.23     0.28       0.8       0.78  <-- within noise
tx_summon                   +1.13     0.29       3.9       1.00
weapon_value                -0.35     0.26       1.3       0.98  <-- within noise
release-gate noise floor (p95 of the bootstrap null): 27.84% of picks flip on resampling alone
wrote D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training\metrics.json
pick flips vs committed: 20.07% of same-class triples (resampling noise floor 27.84%); per-card score volatility p95 16.7 pts
MATERIAL: this refit changes more recommendations than noise would - worth a review.