test(v7): causal variance decomposition of the same-tier spread (C0-C4) + solver utilization audit

Answers the round-3 decision gate question with a controlled experiment on the
official 96 unsolved Lv50 A dataset (shared graph: 768 edges, 8 opponents/card,
same 3 paired Match Seeds per variant, 4,608 mirrored battles each, all BT
converged):

  C0 baseline                         spread 6.642
  C1 scheduling->medians              7.356  (-5.3% variance)
  C2 magnitudes->medians              6.150  (+15.7%)
  C3 both                            6.214  (+21.3%, spread reduction 6.4%)
  C4 scalar-bundle permutation        6.386  (Spearman vs C0 0.801, vs DONOR 0.229)

Decision gate: C3 spread reduction 6.4% <= 40% -> the 'action scalar numbers are
the dominant cause' hypothesis is NOT supported; ~79% of the variance is mechanic
topology. C4's donor correlation of 0.229 is the strongest causal signal that
strength stays with the skeleton, not the drawn numbers. No action-numeric solver
prototype is justified; the Mechanism-Aware StrengthModel remains the path.

Also: solver knob utilization audit over 1000 cards (solver telemetry added in
src/solver-v7.js): MAX_HP first-selected 62% with the largest marginal (it feeds
both endurance and sustain), but 6 of 11 knobs (DEF/RES/SPD/EVA/CRIT/ENERGY_REGEN)
are never selected - a model sensitivity imbalance that any future numeric
extension must fix first.

Official 96-card mechanism ranking under the seed-family split: predictTheta
-0.226, predictReality -0.248, BPv4 -0.059 - all far below the 0.85 gate, so the
Solver/Full-Reality stage stays closed.

Artifacts: qa/v7-attribution-C{0..4}.json, qa/v7-scheduling-magnitude-attribution.json,
qa/v7-mechanism-ranking-96.json, qa/v7-solver-utilization.json; docs:
docs/V7-MECHANISM-ATTRIBUTION.md, delivery report round 3.

Author: NBL Agent

RELEASE-MANIFEST.json

@@ -235,8 +235,13 @@
     },
     {
       "path": "docs/V7-DELIVERY-REPORT.md",
-      "size": 15807,
-      "sha256": "7608a45e5b03afa50745f6a7da15ba1a969572ea1ee1cfdc29d2f268aa1d2721"
+      "size": 21203,
+      "sha256": "0c48bfc8ee99755c183f706d957224d873cf4eb0fde4f5ed847c8cbffc798765"
+    },
+    {
+      "path": "docs/V7-MECHANISM-ATTRIBUTION.md",
+      "size": 7151,
+      "sha256": "1bf2e3d00c5d0971dce9d22a6c4278612d0e5672336a1ea49d253508d578954c"
     },
     {
       "path": "docs/numeric-component-catalog.json",
@@ -300,8 +305,8 @@
     },
     {
       "path": "package.json",
-      "size": 3621,
-      "sha256": "ddab4d80377443beb3b202ae8fe30a91282e62475e9f36c0e94195cec930c81f"
+      "size": 3940,
+      "sha256": "dc8732095b1b1f690b5718edefa99554fd190466106d45c657627e83d36674f7"
     },
     {
       "path": "qa/adjacent-rarity-matrix.json",
@@ -843,6 +848,31 @@
       "size": 21988,
       "sha256": "263ab25c4e22940f1d2e1f4c53b1739126dfc1b7ee837b46c137f06b500fcfda"
     },
+    {
+      "path": "qa/v7-attribution-C0.json",
+      "size": 134808,
+      "sha256": "1e3798197a2d3316df431f461c7e79359b32350e967b26a787c567eda74596a6"
+    },
+    {
+      "path": "qa/v7-attribution-C1.json",
+      "size": 134818,
+      "sha256": "a586832fa0d5ae75c8117b0e527911183cbe9185f3a43005a2bc5aab4666a698"
+    },
+    {
+      "path": "qa/v7-attribution-C2.json",
+      "size": 134819,
+      "sha256": "f96c0e03e1ebb6774584968a508694e5ea05263efb107163fe8a63ff0eb2e3a7"
+    },
+    {
+      "path": "qa/v7-attribution-C3.json",
+      "size": 134792,
+      "sha256": "2b41b110124cc11d6fb4572717d630318d411e18c454d4f5375230bbb092d875"
+    },
+    {
+      "path": "qa/v7-attribution-C4.json",
+      "size": 134801,
+      "sha256": "25bdb094352cfe79374b3be9783e2d320eb29cdefc2638dd8bacde2850a00a51"
+    },
     {
       "path": "qa/v7-axis-sensitivity.json",
       "size": 737,
@@ -873,6 +903,11 @@
       "size": 96844,
       "sha256": "7135d739603b89c2f17da8f7ec5d82438943bab919c0d9678b8bfb22441da4ab"
     },
+    {
+      "path": "qa/v7-mechanism-ranking-96.json",
+      "size": 3845,
+      "sha256": "487451c9ce848d9e7e5cc8f18a0cd2a25e7403fd25505ea292c53933671a4996"
+    },
     {
       "path": "qa/v7-mechanism-ranking.json",
       "size": 32269,
@@ -898,6 +933,11 @@
       "size": 1588,
       "sha256": "54b484922ff31c7275b9c243c7cf411e5e5a04a521bb0d912c92044a22d0c5ea"
     },
+    {
+      "path": "qa/v7-scheduling-magnitude-attribution.json",
+      "size": 14984,
+      "sha256": "5fd35f293ebeff55081fb440509fcfa4c9fc7c97defa409c490dcb7df10ccea0"
+    },
     {
       "path": "qa/v7-seed-dispersion.json",
       "size": 17809,
@@ -908,6 +948,11 @@
       "size": 2968,
       "sha256": "3448d259453b839b49587804fedce200a80d7a8f0dd4c578fb0dab0e826e08cf"
     },
+    {
+      "path": "qa/v7-solver-utilization.json",
+      "size": 6649,
+      "sha256": "f0ef1cbf1383779413bc8d092acfe29374b8f0790db17865473800cc0f288d90"
+    },
     {
       "path": "qa/v7-stat-sensitivity.json",
       "size": 92515,
@@ -1033,6 +1078,11 @@
       "size": 4054,
       "sha256": "5a24505e42ffdf1ebcca748276cbedfb209f3fb989bb4a51182e4c2c2e4904c2"
     },
+    {
+      "path": "scripts/audit-v7-solver-utilization.js",
+      "size": 5029,
+      "sha256": "c8f1805cfdb057919fef2088d11dbcb294519927edb8c3884633803007ef353d"
+    },
     {
       "path": "scripts/benchmark-gen-v7.js",
       "size": 2284,
@@ -1073,6 +1123,16 @@
       "size": 3491,
       "sha256": "3516a02e09cdeab2daaaf1290ac07d1e10f5ebb83eda412038660dff7178ec2b"
     },
+    {
+      "path": "scripts/evaluate-v7-mechanism-ranking-96.js",
+      "size": 6370,
+      "sha256": "97e0ceb1d67e929d09ee1e675844e4790800b59e8e4337436f083b84c1c7da7a"
+    },
+    {
+      "path": "scripts/experiment-v7-attribution.js",
+      "size": 17844,
+      "sha256": "48ec468a9fd9aff1e717b9e250819f9f79705f7b2eadf3c7e6a52694bb940811"
+    },
     {
       "path": "scripts/explore-v7-mechanism-features-v2.js",
       "size": 13129,
@@ -1170,8 +1230,8 @@
     },
     {
       "path": "scripts/static-check.js",
-      "size": 9424,
-      "sha256": "b3871c40ed86120157f1771e4a9a4dc267d9555b7894cb46c620c97e46f55e31"
+      "size": 9803,
+      "sha256": "4706d73efee46082fc6508ab55b98109603ae73008771838602b38ab7ba53822"
     },
     {
       "path": "scripts/stress-check.js",
@@ -1325,8 +1385,8 @@
     },
     {
       "path": "src/gen-v7.js",
-      "size": 9357,
-      "sha256": "e8e1a0197042924ae6658a2de37b8f0922be9387ab3d94731b5bf045a742b5d0"
+      "size": 9384,
+      "sha256": "90798da935b11ec10e85a972018e621a9279fd83e3721c5afae4c1bee9fa7977"
     },
     {
       "path": "src/generator.js",
@@ -1385,8 +1445,8 @@
     },
     {
       "path": "src/solver-v7.js",
-      "size": 3269,
-      "sha256": "a60e823f8fc4ed23915ec427f876cffe8d7f005ff627b3152c3fceb3b15045ca"
+      "size": 4331,
+      "sha256": "9e47df1f5b2cf542b28ccfe5a2a032c26b1dd2bba51b8d4652ade3235d64a0ba"
     },
     {
       "path": "src/status-runtime.js",

docs/V7-DELIVERY-REPORT.md

@@ -9,6 +9,102 @@
 > Every number below comes from an actual run of the committed scripts against
 > the committed artifacts. Nothing is estimated or invented.
 
+## Convergence round 3 — causal variance decomposition (user decision gate)
+
+**The question this round had to answer: is the ~6.6 theta same-tier spread
+"real mechanism value" or "seed randomly drawing strong/weak action numbers"?**
+A deterministic counterfactual experiment on the official 96 unsolved Lv50 A
+dataset (shared graph, 768 edges, 8 opponents/card, same 3 paired Match Seeds per
+variant, 4,608 mirrored battles each, all BT-converged) gives the answer:
+
+| variant | what changed | p95−p5 | variance reduction | Spearman(C0, Ck) |
+|---|---|---|---|---|
+| C0 | baseline | **6.642** | — | — |
+| C1 | cost/cooldown/priority → per-slot medians | **7.356** | −5.3% | 0.920 |
+| C2 | damage/heal/barrier coeffs, accuracy, chance, drainRatio, gain → medians | **6.150** | +15.7% | 0.962 |
+| C3 | C1 + C2 | **6.214** | +21.3% | 0.899 |
+| C4 | scalar bundles deterministically permuted (i ← donor i+1) | **6.386** | +13.4% | 0.801 (donor: **0.229**) |
+
+**Decision gate verdict: C3 spread reduction = 6.4% ≤ 40% → hypothesis not
+supported → continue Mechanism-Aware StrengthModel; no action-numeric solver
+prototype.** Readings:
+
+* Normalising both scheduling and magnitudes leaves ~6.2 theta of 6.64 (≈79% of
+  the variance): the spread is dominated by mechanic topology, not by drawn
+  numbers.
+* C1 (scheduling) *raised* the spread — random scheduling partially equalised
+  classes; it is neither the driver nor a promising knob.
+* C4 is the strongest causal signal: after swapping the scalar bundles, ranking
+  correlates 0.801 with the original cards but only 0.229 with the donor cards —
+  strength stays with the skeleton, not the numbers.
+* Q4: pure mechanic topology still produces **≈6.2 theta** — large, so the
+  StrengthModel must genuinely learn mechanism value.
+
+### Q1 — official 96-card mechanism ranking (three splits)
+
+`qa/v7-mechanism-ranking-96.json`; real spread p95−p5 = 6.642.
+
+| model | overall | seed-family test (official) | random test (reference) | mechanism-family holdout (stress) |
+|---|---|---|---|---|
+| predictTheta | −0.222 | **−0.226** | −0.727 | −0.453 |
+| predictReality (committed ridge) | −0.144 | **−0.248** | +0.064 | −0.262 |
+| BattlePower V4 | −0.183 | **−0.059** | −0.673 | −0.197 |
+
+Spread ratios (predicted/real on test): predictTheta 0.11 (crushed), predictReality
+0.91, BPv4 0.55. **All three models fail the ≥0.85 gate and mostly rank at ~0 or
+negatively on the official seed-family test.** The gate is untouched and the
+Solver/Full-Reality stage stays closed.
+
+### Q5 — why the solver mostly adjusts HP (utilization audit)
+
+`qa/v7-solver-utilization.json`, 1000 cards, 5,525 adjustments (5.5/card):
+
+| knob | adjusted% | first-selected% | adjustments | theta share | marginal median | kept-initial% |
+|---|---|---|---|---|---|---|
+| MAX_HP | 76.3 | **62.1** | 2,005 | **0.379** | 3.20 | 24 |
+| ATK | 65.9 | 0.0 | 1,645 | 0.312 | 3.17 | 34 |
+| ACC | 89.0 | 37.1 | 1,580 | 0.289 | 3.14 | 11 |
+| HEAL_POWER | 23.5 | 0.0 | 247 | 0.018 | 1.23 | 77 |
+| PEN | 4.8 | 0.0 | 48 | 0.002 | 0.64 | 95 |
+| DEF / RES / SPD / EVA / CRIT / ENERGY_REGEN | **0.0** | 0.0 | **0** | 0.000 | — | 100 |
+
+Diagnosis: it is **not** a pure greedy-degeneracy story — MAX_HP's marginal is
+genuinely the largest (it feeds both endurance and sustain, since heal/barrier
+formulas scale with MAX_HP), so picking it first (62%) is rational. But the model
+has a **sensitivity imbalance: 6 of 11 knobs are dead** (DEF/RES/SPD/EVA/CRIT/
+ENERGY_REGEN never selected because their marginals are ~100× smaller than the
+top three and can never win the greedy score), so the iso-power surface is
+effectively driven by three knobs. Any future action-numeric extension must fix
+this imbalance first, otherwise it would just add more dead knobs.
+
+### Q6 — parameter classification
+
+Full table in `docs/V7-MECHANISM-ATTRIBUTION.md`. Summary: damage/heal/barrier
+coefficients, periodic magnitude, drainRatio, effect chance and accuracy are
+continuous, monotonic, strength-only → safe continuous marginal-value knobs; cost
+and cooldown are discrete but monotonic → borderline; **priority is discrete,
+non-monotonic and identity/behavior-affecting (it changes AI action choice and
+initiative order) → NOT a strength knob**; shield/ward kind, control status
+flavour, stacks and all structural fields (family set, effects topology,
+statuses, triggers, targets, damage type, resource) are Frozen Identity.
+
+### Q7 — seed identity invariants (proposal, not approved)
+
+Frozen Identity (never touched): action family set/count, victory path, targets,
+trigger topology, status identity, conditionals, periodic topology, resource
+mechanic, damage type, signature action, shield-vs-ward kind, control flavour.
+Shape-Preserving Numeric Freedom: per-mechanism relative shape kept, only
+family/global scaling allowed. Free Strength Scalars: absolute magnitudes that
+change no relative shape or identity.
+
+### Q8 — recommendation
+
+**A — mainly fix the StrengthModel** (with D-style support: then evaluate a
+bounded, shape-preserving numeric freedom, not free per-action knobs). The C0-C4
+evidence rules out B: scalar numbers explain ≤21% of the variance and the C4
+donor test (0.229) shows strength does not follow the numbers; ~6.2 theta of
+topology-driven spread remains, which only a mechanism-aware model can capture.
+
 ## Convergence round 2 — root causes, fixes, and what still blocks
 
 ### Root Cause Before (measured on the inherited code)

docs/V7-MECHANISM-ATTRIBUTION.md

@@ -0,0 +1,110 @@
+# V7 Mechanism Attribution — Analysis Framework
+
+Status: analysis document (round 3, user decision gate). Experimental numbers are
+filled in from `qa/v7-scheduling-magnitude-attribution.json` when the C0-C4 runs
+complete. This page records the parts that are derivable from source alone.
+
+## 0. Causal variance decomposition — RESULT (96 unsolved Lv50 A cards)
+
+Same shared graph (768 edges, 8 opponents/card) and same paired Match Seeds
+(3/card) across all variants; each variant = 4,608 mirrored canonical-AI battles,
+Bradley-Terry fitted, all converged. Card scalars are normalized to per-slot
+medians; skeleton/effects/statuses/triggers/targets are never changed.
+
+| variant | what changed | p95−p5 | variance reduction | spread reduction | Spearman(C0θ, Ckθ) |
+|---|---|---|---|---|---|
+| C0 | baseline | **6.642** | — | — | — |
+| C1 | cost/cooldown/priority → medians | **7.356** | **−5.3%** | −10.8% | 0.920 |
+| C2 | damage/heal/barrier coeffs, accuracy, chance, drainRatio, gain → medians | **6.150** | **+15.7%** | +7.4% | 0.962 |
+| C3 | C1 + C2 | **6.214** | **+21.3%** | **+6.4%** | 0.899 |
+| C4 | deterministic permutation of scalar bundles across cards (card i ← donor i+1) | **6.386** | +13.4% | +3.9% | 0.801 (vs donor C0θ: **0.229**) |
+
+**Decision gate (user §4): C3 spread reduction = 6.4% ≤ 40% → "hypothesis not
+supported; continue Mechanism-Aware StrengthModel."** An action-numeric solver
+prototype is therefore NOT justified by this evidence.
+
+Readings:
+
+* Normalizing BOTH random scheduling and scalar magnitudes leaves ~6.2 theta of
+  the original 6.64 — roughly **79% of the variance** survives as
+  topology/interactions. The same-tier spread is dominated by the mechanic
+  skeleton (family, effects, statuses, triggers, targets, sustain-vs-damage
+  races), not by the drawn numbers.
+* C1 *increased* the spread (−5.3% variance): the random scheduling partially
+  equalised mechanism classes; removing it made cards more different. Scheduling
+  randomness is not the driver and is not a promising solver knob.
+* C2 (magnitude scalars) is the largest single contributor (+15.7%) but still
+  modest.
+* C4 is the strongest causal signal: after permuting the scalar bundles, the new
+  ranking correlates 0.801 with the ORIGINAL cards' ranking but only **0.229**
+  with the DONOR cards' ranking — i.e. strength stays with the skeleton and does
+  NOT follow the numbers.
+* Q4 answer: pure mechanic topology after normalization still produces **≈6.2
+  theta** of spread — large. Mechanisms genuinely differ hugely in long-grind
+  value; the StrengthModel must learn to value them (it currently ranks them at
+  ≈0 or worse, see `qa/v7-mechanism-ranking-96.json`).
+
+## 1. What the V7 skeleton draws randomly vs structurally (from `src/gen-v7.js`)
+
+Every non-backbone action is constructed with
+`cost=1+random(5), cooldown=1+random(4), priority=random(5)-2` from ONE shared
+PRNG stream, and each family case may override some fields:
+
+| action (family) | cost | cooldown | priority | other drawn scalars |
+|---|---|---|---|---|
+| 基础攻势 backbone | 0 fixed | 1 fixed | 10 fixed | status chance `0.55+0.3·triggers` (genome) |
+| 精确打击 direct | random 1-5 | random 1-4 | random -2..2 | dmg `24+24·pressure` (genome) |
+| 裂变突袭 burst | random | 3 fixed | random | dmg `16+16·pressure` ×2 hits (genome) |
+| 蚀印扩散 dot | random | random | random | stacks `1+random(2)` (RANDOM), chance `0.65+0.3·reliability` |
+| 汲取 drain | random | random | random | dmg `20+20·pressure`, drainRatio `0.08+0.22·sustain` |
+| 压制 control | 0 fixed | 2 fixed | random | dmg `14+14·pressure`, status pick (RANDOM), chance `0.6+0.3·control` |
+| 屏障 barrier | 0 fixed | 2 fixed | random | type pick shield/ward (RANDOM), coeff `0.58+0.24·endurance` |
+| 再生 sustain | 0 fixed | 2 fixed | random | heal coeff `0.62+0.24·sustain` |
+| 蓄能轰击 resource | random | random | random | gain `2+random(3)` (RANDOM), dmg `18+20·pressure` |
+| 先制 tempo | random | 1 fixed | `2+3·tempo` fixed | dmg `18+20·pressure` |
+| 反射架势 reflect | random | random | random | (status only) |
+
+`accuracy = 0.72+0.24·reliability` for every non-backbone action (genome). The
+anchor `55·(ATK+1)·(0.02+ACC/100)·(0.70+PEN/100)` is prepended to every action
+(structural). Resource type, damage type, trigger event, victory path, action
+family set, statuses and durations all come from the seed-driven skeleton and are
+structural.
+
+## 2. Parameter classification (Q6)
+
+| parameter | continuous/discrete | monotonic in strength | strength-only / identity-affecting | safe as solver knob? |
+|---|---|---|---|---|
+| damage coefficient (`ATK * X`) | continuous | monotonic ↑ | strength-only | YES (continuous marginal-value knob) |
+| heal coefficient (`MAX_HP * X`) | continuous | monotonic ↑ | strength/sustain | YES |
+| barrier coefficient | continuous | monotonic ↑ | strength/durability | YES |
+| periodic magnitude (`ATK * X`) | continuous | monotonic ↑ | strength (dot) | YES |
+| drainRatio | continuous | monotonic ↑ | strength/sustain | YES |
+| effect chance | continuous | monotonic ↑ | strength/delivery | YES (bounded 0..1) |
+| action accuracy | continuous | monotonic ↑ | strength/delivery | YES |
+| energy cost | discrete int | monotonic ↓ | strength/availability | borderline — discrete, but monotonic |
+| cooldown | discrete int | monotonic ↓ | strength/availability | borderline — discrete, but monotonic |
+| priority | discrete int | NON-monotonic (order, not magnitude) | **identity/behavior** (changes AI action choice and initiative order) | NO as plain strength knob |
+| shield-vs-ward type | categorical | — | identity | NO |
+| stun/slow/silence pick | categorical | — | identity (control flavour) | NO |
+| dot stacks | discrete int | monotonic ↑ | strength (dot magnitude) | borderline |
+| action family / target / effect topology / statuses / triggers / damage type / resource | structural | — | **Frozen Identity** | NEVER |
+
+## 3. Seed Identity Invariants (Q7)
+
+Proposed split for any future action-numeric prototype:
+
+* **Frozen Identity** (solver never touches): action family set and count,
+  victory path, target types, trigger topology, status/debuff identity,
+  conditional structure, periodic topology, resource mechanic, damage type,
+  signature action identity, shield-vs-ward kind, control status flavour.
+* **Shape-Preserving Numeric Freedom**: per-mechanism relative shape (e.g. the
+  ratios between a kit's own action magnitudes, its cooldown/resource tendency,
+  its sustain tendency) stays as the seed drew it; the solver may only apply
+  FAMILY/GLOBAL scaling factors so General Strength reaches TargetTheta.
+* **Free Strength Scalars**: absolute magnitudes that do not change any relative
+  shape or identity (global coefficients, global availability level), free for
+  iso-power convergence.
+
+The three-way split is a candidate architecture, not yet approved; this round
+only produces the causal evidence (C0-C4) and the solver-utilization audit that
+inform whether an action-numeric prototype is justified at all.

package.json

@@ -12,6 +12,9 @@
     "audit:v7-solver-saturation": "node scripts/audit-v7-solver-saturation.js 1000",
     "diagnose:v7-mechanism": "node scripts/diagnose-v7-mechanism-ranking.js 48 10 3",
     "explore:v7-mechanism": "node scripts/explore-v7-mechanism-features-v2.js",
+    "experiment:v7-attribution": "node scripts/experiment-v7-attribution.js run C3 96 8 3 && node scripts/experiment-v7-attribution.js aggregate",
+    "audit:v7-solver-utilization": "node scripts/audit-v7-solver-utilization.js 1000",
+    "evaluate:v7-mechanism-96": "node scripts/evaluate-v7-mechanism-ranking-96.js",
     "gate:v6-strength": "node scripts/gate-v6-strength.js",
     "reality:v7": "node scripts/build-v7-battle-graph.js 15 12 100 && node scripts/fit-v7-empirical-strength.js && node scripts/calibrate-strength-model-v7.js && node scripts/audit-v7-seed-dispersion.js && node scripts/audit-v7-product-strength.js 40 8 && node scripts/audit-stat-sensitivity-v7.js 7 8 8 3 && node scripts/audit-battlepower-v4.js",
     "build:v7-graph": "node scripts/build-v7-battle-graph.js 15 12 100",
@@ -49,3 +52,4 @@
   }
 }
 
+