Retrain heuristic weights (2026-08-14)

[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: 2208
rows: 2208 (single source)
features: 31 fitted, 19 dropped
  dropped: kw_charge(n=11), kw_colossal(n=11), kw_combo(n=21), kw_echo(n=5), kw_forge(n=3), kw_freeze(n=23), kw_magnetic(n=13), kw_outcast(n=18), kw_poisonous(n=37), kw_reborn(n=39), kw_spellpower(n=18), kw_stealth(n=13), kw_windfury(n=8), tx_armor(n=28), tx_destroy_minion(n=28), tx_discover(redundant), tx_restore_amt(n=18), tx_silence(n=17), tx_transform(n=34)
penalty BEFORE normalizing weights: alpha=10 vs mean(diag)=77471.3  -> shrinkage 0.000129, df_eff=30.0/31
CV: 5-fold grouped by card (1093 cards, 2208 rows), 5 repeats, 11 alphas

     alpha   withinClassRho       se
      0.03           0.1733   0.0016
       0.1           0.1733   0.0016
       0.3           0.1735   0.0016
         1           0.1740   0.0016
         3           0.1755   0.0014
        10           0.1793   0.0011
        30           0.1844   0.0009
       100           0.1835   0.0012
       300           0.1748   0.0013
      1000           0.1554   0.0025
      3000           0.1289   0.0023
CV best alpha=30 (rho=0.1844); 1-SE rule picks alpha=30 (rho=0.1844, threshold 0.1835)
penalty AFTER: alpha=30 vs mean(diag)=2396.5  -> shrinkage 0.012364, df_eff=28.9/31

holdout evaluation (pooled Spearman on held-out rows, alpha=30):
  random cards (baseline)            n= 109  rho= 0.2271  (se 0.0346)
  lowest-games decile                n= 109  rho= 0.2463  (up to 219 games)
  -> covariate-shift gap vs baseline: +0.0192
  target spread / model error:
      thin  sd(y)=6.582  MAE=5.190
      rest  sd(y)=4.986  MAE=3.210
  calibration slope (truth regressed on prediction):
      thin decile  1.2622   random holdout  0.9544  (se 0.1101)
      -> measured ModelOnlyShrink = 1.0000 (thin/random; set the runtime constant from THIS, not from a correlation ratio)
  leave-out set CORE                 n= 210  rho= 0.1539
  leave-out set TIME_TRAVEL          n= 171  rho= 0.2879
  leave-out set BATTLE_OF_THE_BANDS  n= 158  rho= 0.3495
  leave-out set WHIZBANGS_WORKSHOP   n= 151  rho= 0.2435

fitting ridge (alpha=30) on 2208 rows, 31 features...
display anchor (pool median raw): -0.97, robust sigma: 1.2466 -> 15 pts per robust SD
wrote D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training\arena_weights.generated.json

feature                  committed   generated     diff
intercept                   +1.09      +1.05    -0.04
anchor_median_raw           -0.92      -0.97    -0.05  <-- differs
attack                      -0.15      -0.12    +0.03  <-- differs
cost                        +0.15      +0.08    -0.07  <-- differs
has_tribe                   +0.12      +0.39    +0.27  <-- differs
health                      -0.18      -0.12    +0.06  <-- differs
is_hero                     +0.81      +3.83    +3.02  <-- differs
is_loc                      -0.88      -0.77    +0.11  <-- differs
is_minion                   +2.61      +2.42    -0.19  <-- differs
is_neutral                  -1.94      -2.08    -0.14  <-- differs
is_spell                    -2.75      -2.53    +0.22  <-- differs
is_weapon                   -0.94      -1.22    -0.28  <-- differs
kw_battlecry                +0.70      +0.60    -0.10  <-- differs
kw_deathrattle              -0.41      -0.74    -0.33  <-- differs
kw_discover                 +0.86      +0.62    -0.24  <-- differs
kw_divine_shield            +0.10      +0.22    +0.12  <-- differs
kw_lifesteal                +0.84      +0.89    +0.05  <-- differs
kw_reborn                   +1.83      +0.00    -1.83  <-- differs
kw_rush                     +1.05      +1.16    +0.11  <-- differs
kw_secret                   -0.19      -0.07    +0.12  <-- differs
kw_taunt                    -0.10      +0.00    +0.10  <-- differs
kw_tradeable                +0.71      +0.78    +0.07  <-- differs
stat_per_mana               -2.40      -2.31    +0.09  <-- differs
statline                    +0.34      +0.36    +0.02  <-- differs
tx_aoe                      +0.51      +0.43    -0.08  <-- differs
tx_damage_amt               +0.08      +0.08    +0.00
tx_dmg_per_mana             +0.00      -0.07    -0.07  <-- differs
tx_draw                     -0.31      -0.26    +0.05  <-- differs
tx_gain_card                -0.87      -0.49    +0.38  <-- differs
tx_mana_cheat               +1.05      +0.75    -0.30  <-- differs
tx_persistent               +0.15      +0.23    +0.08  <-- differs
tx_random                   +0.32      +0.13    -0.19  <-- differs
tx_summon                   +0.95      +1.38    +0.43  <-- differs
weapon_value                -0.20      -0.28    -0.08  <-- differs

max abs weight diff: 3.02
weights differing by >= 0.01: 32
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, 22.57 },
  { HSCard.ElvenArcher, 52.11 },
  { HSCard.DefenderOfArgus, 39.15 },
  { HSCard.FieryWarAxe, 50.96 },
  { HSCard.Fireball, 52.48 },
  { HSCard.ColdlightOracle, 43.50 },
  { HSCard.DireFrenzy, 47.71 },
  { HSCard.NorthshireCleric, 46.99 },
  { HSCard.DarkIronDwarf, 29.11 },
  { HSCard.Deathwing, 17.63 },

bootstrapping 300 card-resampled refits...

pick instability under resampling, by how clearly the model named a winner:
  winner leads by >  0.0 pts:  27.58% of those picks flip
  winner leads by >  1.0 pts:  25.69% of those picks flip
  winner leads by >  3.0 pts:  22.23% of those picks flip
  winner leads by >  5.0 pts:  19.18% of those picks flip
  winner leads by > 10.0 pts:  13.14% of those picks flip

feature                    weight       se    |w|/se   signCons
attack                      -0.12     0.11       1.1       0.90  <-- within noise
cost                        +0.08     0.12       0.7       0.74  <-- within noise
has_tribe                   +0.39     0.25       1.5       0.93  <-- within noise
health                      -0.12     0.13       0.9       0.81  <-- within noise
is_hero                     +3.83     4.45       0.9       0.68  <-- within noise
is_loc                      -0.77     1.46       0.5       0.66  <-- within noise
is_minion                   +2.42     1.18       2.0       1.00
is_neutral                  -2.08     0.32       6.6       1.00
is_spell                    -2.53     1.04       2.4       1.00
is_weapon                   -1.22     1.54       0.8       0.71  <-- within noise
kw_battlecry                +0.60     0.29       2.0       0.99
kw_deathrattle              -0.74     0.31       2.4       0.99
kw_discover                 +0.62     0.33       1.9       0.96  <-- within noise
kw_divine_shield            +0.22     0.53       0.4       0.69  <-- within noise
kw_lifesteal                +0.89     0.45       2.0       0.97  <-- within noise
kw_rush                     +1.16     0.41       2.8       1.00
kw_secret                   -0.07     0.55       0.1       0.60  <-- within noise
kw_taunt                    -0.04     0.39       0.1       0.54  <-- within noise
kw_tradeable                +0.78     0.66       1.2       0.84  <-- within noise
stat_per_mana               -2.31     1.64       1.4       0.85  <-- within noise
statline                    +0.36     0.14       2.6       0.99
tx_aoe                      +0.43     0.38       1.1       0.89  <-- within noise
tx_damage_amt               +0.08     0.15       0.5       0.68  <-- within noise
tx_dmg_per_mana             -0.07     0.63       0.1       0.54  <-- within noise
tx_draw                     -0.26     0.41       0.6       0.72  <-- within noise
tx_gain_card                -0.49     0.34       1.4       0.94  <-- within noise
tx_mana_cheat               +0.75     0.42       1.8       0.97  <-- within noise
tx_persistent               +0.23     0.36       0.7       0.70  <-- within noise
tx_random                   +0.13     0.30       0.4       0.73  <-- within noise
tx_summon                   +1.38     0.33       4.2       1.00
weapon_value                -0.28     0.26       1.1       0.90  <-- within noise
release-gate noise floor (p95 of the bootstrap null): 27.58% of picks flip on resampling alone
wrote D:\a\HdtArenaHelper\HdtArenaHelper\HdtArenaHelper.Training\metrics.json
pick flips vs committed: 16.94% of same-class triples (resampling noise floor 27.58%); per-card score volatility p95 17.1 pts
MATERIAL: this refit changes more recommendations than noise would - worth a review.