add-ai-feature-policy

name	add-ai-feature-policy
description	Use when adding a new deck-aware AI feature (`DeckFeatures` axis) and its companion policies (`TacticalPolicy` + optional `MulliganPolicy`) to `phase-ai`. Codifies the structural-detection pattern, AST-type locations, parts-based predicate convention, registry wiring sites, and recurring review traps so a new feature ships in a single commit without re-research.

Adding a Deck-Aware AI Feature + Policy

This is the authoritative checklist for adding a new feature to the three-layer AI architecture (Features → Plan → Policies). Nine features exist as reference: landfall, mana_ramp, tribal, control, aristocrats, aggro_pressure, tokens_wide, plus_one_counters, spellslinger_prowess. Each follows the same pattern; mirror it.

Before you start: Read landfall.rs (simplest), then aristocrats.rs (geometric-mean commitment with identity lookup), then one feature similar in shape to yours. Trace from features/<name>.rs::detect() through DeckFeatures::analyze and AiSession::from_game through policies/<name>.rs::verdict so the data flow is concrete.

CR Verification Rule. Every CR number in annotations MUST be verified by grepping docs/MagicCompRules.txt BEFORE you write it. Do NOT rely on memory or prior prompts. The 701.x and 702.x sections are arbitrary sequential assignments that hallucinate consistently. Run grep -n "^701.34" docs/MagicCompRules.txt for every number. CR 122.1d is Stun (NOT +1/+1 — that's CR 122.1a). CR 701.27 is Transform (NOT Proliferate — that's CR 701.34). Two real catches from the existing batch — you will have similar misses unless you grep.

---

1. Architecture (memorize this)

AiSession (per-game cache; built once in AiSession::from_game)
  ├─ features: HashMap<PlayerId, DeckFeatures>    ← Layer 1: structural detection
  ├─ plan:     HashMap<PlayerId, PlanSnapshot>    ← Layer 2: bottoming + curve schedule
  ├─ deck_profile / strategy: per-player strategic profiles
  ├─ synergy:  HashMap<PlayerId, SynergyGraph>
  └─ memory:   PolicyMemory

DeckFeatures
  ├─ landfall:       LandfallFeature
  ├─ mana_ramp:      ManaRampFeature
  ├─ tribal:         TribalFeature
  ├─ control:        ControlFeature
  ├─ aristocrats:    AristocratsFeature
  ├─ aggro_pressure: AggroPressureFeature
  ├─ tokens_wide:    TokensWideFeature
  ├─ plus_one_counters: PlusOneCountersFeature
  ├─ spellslinger_prowess: SpellslingerProwessFeature
  └─ <your new feature>

Policies (Layer 3)
  ├─ TacticalPolicy — in-game decisions (CastSpell, ActivateAbility, DeclareAttackers, ...)
  └─ MulliganPolicy — pre-game hand evaluation (sibling trait)

Both run inside their respective registries (`PolicyRegistry::default`,
`MulliganRegistry::default`). Activation is the single multiplicative knob.

AiSession::from_game analyzes current_main plus current_commander; commander entries are weighted as build-around cards before feature/profile/synergy detection. One-off analysis helpers that only receive a deck slice stay pure over that slice.

Single-knob rule. A TacticalPolicy exposes activation() returning Option<f32>. The registry multiplies verdict.delta * activation exactly once. There is no archetype_scale, no turn_mult, no second pass. If a policy needs archetype/turn-sensitive weight, it computes it inside activation() from the inputs.

Score contract. Tactical policy deltas are card-equivalent units: delta = 1.0 means one card of expected value. Use the band helpers on PolicyVerdict so the chosen strength is visible at the call site:

Band	Helper	Range	Meaning
Nudge	PolicyVerdict::nudge	(0.0, 0.3]	tie-breaker
Preference	PolicyVerdict::preference	(0.3, 1.5]	half-card to 1.5-card preference
Strong	PolicyVerdict::strong	(1.5, 5.0]	multi-card swing
Critical	PolicyVerdict::critical	(5.0, 15.0]	game-deciding
Veto	PolicyVerdict::reject	n/a	never take this action

All scoring scalars come from AiConfig::policy_penalties / PolicyPenalties with rationale comments. Thresholds that only gate activation stay as pub const in the policy/feature module. Never encode vetoes as sentinel scores.

---

2. Hard constraints (non-negotiable)

Rule	Why	How to apply
Structural detection only. No card-name matching for classification.	Name lists rot; structural patterns generalize across sets.	Walk CardFace.{abilities,triggers,static_abilities,replacements,keywords,card_type}. Use crate::ability_chain::collect_chain_effects for ability chains.
Identity lookup is exempt. Vec<String>::iter().any(|n| n == &obj.name) is allowed when the policy needs to re-find a structurally-classified card on the battlefield or in hand.	Mulligan policies see GameObject shells whose abilities aren't resolved yet.	Populate payoff_names: Vec<String> once per UNIQUE face in detect(), NOT per playset copy. The no_name_matching lint at features/tests/no_name_matching.rs enforces the forbidden-pattern list and exempts identity lookup.
CR annotations grep-verified. Every CR comment must match docs/MagicCompRules.txt.	Hallucinated CR numbers create false confidence.	Run grep -n "^XXX.Y" docs/MagicCompRules.txt for every annotation BEFORE writing. Don't annotate plumbing/utility code.
Single-authority cost rule. Always ability.cost_categories(), never destructure AbilityCost.	Scattered cost handling means new cost types break in a dozen places.	When you need "this ability sacrifices a permanent", use cost_categories().contains(&CostCategory::SacrificesPermanent).
No bool fields. Use existing typed enums (ControllerRef, Comparator, TurnOrder, etc.).	Booleans aren't composable; an enum is self-documenting.	If you need a "your-vs-opponent" distinction, use ControllerRef::You/Opponent. If you need on-play vs on-draw, use the existing TurnOrder.
Use crate::ability_chain::collect_chain_effects.	Single source of truth for ability-chain walking.	Never re-implement collect_chain_effects. If you need to walk a ResolvedAbility chain in a policy, use (handles chain).

---

Performance — verdict() runs in the AI search inner loop

PolicyRegistry::verdicts() (registry.rs:375-433) runs every policy whose decision_kinds() matches the candidate's DecisionKind, calling activation() then verdict(), for every candidate action, at every node the AI search expands. On a large late-game board that is thousands of verdict() calls per decision. Treat verdict() and everything it transitively calls as an inner-loop function — the same discipline you would apply to code inside the engine's search.

The trap that shipped (x_cast_gate, commit 3de827350d). XCastGatePolicy::verdict() calls engine::game::max_x_value per candidate. max_x_value (casting_costs.rs:7592) walks the whole battlefield calling feasible_mana_capacity per permanent, and for a spell adds a per-X concrete_cost_for_x cost-orchestration loop — a board-wide affordability sweep. Run per candidate per search node on a turn-41 board, it regressed the Court of Grace decision from ~2.1s to ~5.7s (isolated A/B). The policy was rules-correct and passed cargo ai-gate with 0 win-rate flips; the entire cost was wall-clock.

Rules:

1. Order predicates cheap-to-expensive; short-circuit on card-local AST before anything board-wide. Check the candidate's own ManaCost / AbilityCost / effect chain (a handful of matches! over one card's AST) first, and reach for a board-wide or affordability engine call only after the card-local checks prove the candidate is even a member of the class you gate. x_cast_gate correctly gates max_x_value behind an {X}-shard check — but any predicate that ANDs a board-wide call with a card-local one must run the card-local one FIRST. AND is commutative, so the ordering is free latency. (For x_cast_gate specifically, the card-local no_op_at_x_zero payoff walk should precede the board-wide max_x_value — it spares the sweep for every {X} card whose payoff has a fixed residual or does not scale with X.)

2. These engine calls are board-wide or state-cloning — never call them unguarded in activation()/verdict(): max_x_value / feasible_mana_capacity (affordability sweep), find_legal_targets (self-heals to a full scan), any SimulationFilter-cloning path, and mana-availability sweeps (the pass-priority mana clone storm and declare-attackers mana sweep are documented quadratic hot paths). Prefer a *_for_simulation variant where one exists. If you must call one, gate it behind the cheapest structural discriminator and push it as late in the predicate as the logic allows.

3. activation() is the free opt-out; use it. It receives only (features, state, player) — not the candidate — so it cannot inspect the specific card, but it CAN switch the whole policy off for decks/states the policy never fires on (if features.<feat>.commitment < FLOOR { None }). A policy that returns Some(1.0) unconditionally (an activation-constant backstop, like x_cast_gate) pays full verdict() cost on every matching candidate — acceptable only when verdict() itself short-circuits cheaply per rule 1.

4. cargo ai-gate is blind to latency, and cargo ai-perf-gate is blind to un-instrumented calls. ai-gate measures win-rate only. () field-wise sums engine fields () against — but it only sees paths that bump a counter. / bump none, so BOTH gates missed the regression above; a manual wall-clock A/B caught it. Therefore, if your policy adds a board-wide/affordability engine call that is not already counter-instrumented, EITHER add a field for it (and refresh the perf baseline) so can see it, OR run and attach a wall-clock A/B on a large late-game board showing no material regression. A 0-flip is necessary but NOT sufficient.

---

3. AST type table — where things live

This table is the single biggest research-saver. Every column has been verified against engine source during the 9-feature batch.

Concept	AST type	File:line
Card face (the deck-time entity)	CardFace { name, card_type, mana_cost, abilities, triggers, static_abilities, replacements, keywords, power, toughness, ... }	crates/engine/src/types/card.rs:30+
Core type set	CardFace.card_type.core_types: Vec<CoreType>; CoreType::{Creature, Instant, Sorcery, Land, Artifact, Enchantment, Planeswalker, Battle, Tribal, Kindred}	card_type.rs:74-77
Mana cost / value	CardFace.mana_cost: ManaCost; ManaCost::mana_value() -> u32	card.rs:51; mana.rs:468
Face-level keywords	CardFace.keywords: Vec<Keyword> (including Keyword::Prowess, Keyword::Haste, Keyword::EtbCounter { counter_type, count }, etc.)	card.rs:60; keywords.rs:24+
Ability definition	CardFace.abilities: Vec<AbilityDefinition>; AbilityDefinition.kind: AbilityKind::{Spell, Activated, Triggered, Static}	card.rs:62; ability.rs:3749
Ability cost	AbilityDefinition.cost: Option<AbilityCost>. Use ability.cost_categories() (returns Vec<CostCategory>) — never destructure variants.	ability.rs:1748+
Effect chain	AbilityDefinition.effect: Effect + sub_ability: Option<Box<AbilityDefinition>>. Use crate::ability_chain::collect_chain_effects(ability) to walk.	ability.rs:2080+
Triggered ability	CardFace.triggers: Vec<TriggerDefinition>; key fields: mode: TriggerMode, valid_card: Option<TargetFilter>, valid_target: Option<TargetFilter>, counter_filter: Option<CounterTriggerFilter>, constraint: Option<TriggerConstraint>,

GameObject vs CardFace API drift. This bites every feature. CardFace has static_abilities/replacements; GameObject has static_definitions/replacement_definitions. Detection (CardFace) and policy (GameObject) need parallel access — use parts-based predicates (Section 5).

---

4. Required artifacts per feature

Every feature must produce ALL of these:

1. AST verification table in module docstring — one line per axis citing the exact crates/engine/src/types/...:line where the type lives + a CR number where applicable. Mirror landfall.rs:1-12 and aggro_pressure.rs:1-25. This is your proof that the AST supports detection without parser changes.
2. <Name>Feature struct — counters as u32, ratios as f32, identity-lookup lists as Vec<String>, single commitment: f32 (0.0..=1.0). NO bool fields. Document each field with a CR if it implements a rule.
3. Commitment formula with calibration anchor in a doc comment. Commitments are density-normalized per 60 nonland cards so 60-card and 99-card decks with the same density score the same. Use only the shared helpers in features/commitment.rs: commitment::density_per_60(count, total_nonland), commitment::weighted_sum(&[(weight, density)]), and commitment::geometric_mean(&[pillar_density]). Show the math AND a real-deck calibration ("Mono-Red Burn: ~16 one-drops + 4 burn → commitment ≈ 0.85"). Anti-calibration ("UW control → commitment ≈ 0.0") is also required. Two formula shapes are established:
   • Weighted sum, clamped (when missing pillars are tolerable). Pattern: control, aggro_pressure, spellslinger.
   • Geometric mean with zero-pillar collapse (when missing a pillar = not this archetype). Pattern: aristocrats, tokens_wide.
4. Parts-based predicate exports — one pub(crate) fn is_<X>_parts(<minimal CardFace slices>) -> bool per axis. Each takes the slice the policy actually has access to on GameObject (e.g., is_burn_spell_parts(core_types: &[CoreType], abilities: &[AbilityDefinition])). Add a public is_X(face: &CardFace) wrapper if useful for tests; remove it if dead per clippy. The _parts predicates are the contract — they let the policy classify live GameObjects without reconstructing a CardFace.
5. Threshold constants — pub const. Mirror tribal.rs:47-67: LORD_PRIORITY_FLOOR, MULLIGAN_FLOOR, AGGRO_TEMPO_FLOOR. Each has a docstring explaining the semantic level it gates.
6. — one CR per rule-bearing comment, every number grep-verified. Don't annotate utility/plumbing code.

---

5. The parts-based predicate convention

This is the architectural heart. Every classifier exists in two shapes:

// Public-facing convenience (used in detect() and tests):
pub fn is_burn_spell(face: &CardFace) -> bool {
    is_burn_spell_parts(&face.card_type.core_types, &face.abilities)
}

// Reusable contract (used in policies against GameObject fields):
pub(crate) fn is_burn_spell_parts(
    core_types: &[CoreType],
    abilities: &[AbilityDefinition],
) -> bool {
    abilities.iter().any(|a| {
        a.kind == AbilityKind::Spell
            && collect_chain_effects(a)
                .iter()
                .any(|e| matches!(e, Effect::DealDamage { target, .. }
                                  if filter_can_target_player(target)))
    })
}

Why parts-based:

• Detection runs against CardFace (the deck list).
• Policies run against GameObject (the live battlefield/hand/stack).
• The fields between them have different names (static_abilities vs static_definitions, replacements vs replacement_definitions) but the same shape.
• Parts predicates take the slice — both call sites pass their respective field. No conversion, no duplication.

Rule of thumb: if your detect() uses face.X, the corresponding parts predicate takes X: &[T] and the policy passes &obj.X (or the equivalently-named runtime field).

---

6. The TacticalPolicy / MulliganPolicy contract

// crates/phase-ai/src/policies/registry.rs
pub trait TacticalPolicy: Send + Sync {
    fn id(&self) -> PolicyId;
    fn decision_kinds(&self) -> &'static [DecisionKind];
    fn activation(&self, features: &DeckFeatures, state: &GameState, player: PlayerId) -> Option<f32>;
    fn verdict(&self, ctx: &PolicyContext<'_>) -> PolicyVerdict;
}

// crates/phase-ai/src/policies/mulligan/mod.rs
pub trait MulliganPolicy: Send + Sync {
    fn id(&self) -> PolicyId;
    fn evaluate(
        &self,
        hand: &[ObjectId],
        state: &GameState,
        features: &DeckFeatures,
        plan: &PlanSnapshot,
        turn_order: TurnOrder,
        mulligans_taken: u8,
    ) -> MulliganScore;
}

activation() patterns:

• if features.<feat>.commitment < FLOOR { None } else { Some(features.<feat>.commitment) } — opt-out gate + linear scaling. Standard.
• Constant Some(1.0) requires an // activation-constant: <reason> marker (lint at policies/tests/activation_marker_lint.rs).

verdict() patterns:

• Return PolicyVerdict::{nudge,preference,strong,critical}(ctx.config.policy_penalties.<named_field>, reason) for scores. Return PolicyVerdict::reject(reason) only for genuine vetoes.
• PolicyReason::new("stable_kind_string").with_fact("metric_name", value as i64) — kind is &'static str, frozen per policy.
• Mirror free_outlet_activation.rs:80-94: re-classify the live ability structurally using parts predicates against obj.abilities/obj.keywords/etc. — don't trust deck-time classification at decision time.

MulliganPolicy::evaluate patterns:

• Opt-out below MULLIGAN_FLOOR returns Score { delta: 0.0, reason: <name>_keepables_na }.
• Walk hand (Vec<ObjectId>); look up each via state.objects.get(&oid). Classify by core type, mana value, identity lookup against feature payoff_names.
• Verdict tiers in priority order (first match returns): _keepable_ideal (+1.5..2.0), _workable (+0.5..1.0), penalty (-1.0..-1.5), default (0.0).
• turn_order and mulligans_taken are real inputs. New policies must consume them or carry an // input-unused: <reason> marker next to the binding.

---

7. Wiring sites (file:line — read each before editing)

For a feature named <feat> with <Feat>Feature struct and policies <Feat>Policy + <Feat>Mulligan:

File	Edit
crates/phase-ai/src/features/mod.rs	Add pub mod <feat>; + pub use <feat>::<Feat>Feature; + pub <feat>: <Feat>Feature field on DeckFeatures.
crates/phase-ai/src/features/<feat>.rs	NEW. The feature module + tests.
crates/phase-ai/src/features/mod.rs	Call <feat>::detect(deck) in DeckFeatures::analyze.
crates/phase-ai/src/policies/mod.rs	Add pub mod <feat>; (and any new tactical policy module).
crates/phase-ai/src/policies/<feat>.rs	NEW. Tactical policy + tests.
crates/phase-ai/src/policies/registry.rs	Add PolicyId::<Feat> (and <Feat>Mulligan) variants. Import super::<feat>::<Feat>Policy. Add Box::new(<Feat>Policy) to PolicyRegistry::default's vec.
crates/phase-ai/src/policies/mulligan/<feat>_keepables.rs	NEW (if a mulligan policy is in scope).
crates/phase-ai/src/policies/mulligan/mod.rs	Add module + pub use + Box::new(<Feat>Mulligan) to MulliganRegistry::default's vec.
crates/phase-ai/src/config.rs	Add the policy's scoring scalar(s) to PolicyPenalties, and register each new field in ACTIVE_POLICY_PENALTY_FIELDS or UNTUNED_POLICY_PENALTY_FIELDS. Not optional — a test enforces it. See Section 7a.
crates/phase-ai/src/plan/curves.rs	OPTIONAL: add tempo-class branch in tempo_class_for (carefully ordered — see Section 8). Add expected-curve adjustments in expected_lands_for/expected_mana_for/expected_threats_for if archetype shifts the curve. Plan data is consumed by mulligan bottoming and curve expectations; do not add zombie plan fields.

The no_name_matching lint at features/tests/no_name_matching.rs automatically scans both src/features/ and src/policies/ — no manual exemption needed unless your file names appear in the lint's allow-list (unlikely).

---

7a. PolicyPenalties registration — ACTIVE vs UNTUNED

Every scoring scalar you add to PolicyPenalties must be listed in exactly one of two sets in crates/phase-ai/src/config.rs. This is enforced, not advisory:

• every_policy_penalty_is_tuning_registered_or_explicitly_untuned (config.rs:1342) serializes PolicyPenalties::default() and asserts set equality between its field names and ACTIVE_POLICY_PENALTY_FIELDS ∪ UNTUNED_POLICY_PENALTY_FIELDS. A new field in neither set — or a stale name in either set — turns cargo nextest run -p phase-ai red (Tilt test-ai; CI runs it workspace-wide via the cargo nextest run --profile ci … --workspace step in .github/workflows/ci.yml). The same test requires every UNTUNED_ entry's reason string to be non-empty.

Which set? ACTIVE_POLICY_PENALTY_FIELDS (config.rs:617) is the CMA-ES parameter vector: ai_tune builds the --group penalties parameter names from it (bin/ai_tune.rs:102) and policy_penalties_from_params deserializes the optimizer's output back over exactly those fields, clamped to [-15.0, 15.0] (bin/ai_tune.rs:222). Anything listed there is a free variable the optimizer will overwrite.

• ACTIVE_ — soft heuristics whose magnitude is a preference (nudge/preference/strong band). Only after a paired-seed cargo ai-gate calibration; that's what every current UNTUNED_ reason is waiting for.
• UNTUNED_ (config.rs:658, (field, reason) pairs) — the default, and the mandatory home for any game-deciding scalar whose value is load-bearing for correctness rather than taste (a critical-band win-detector weight, e.g. a CR 104.2a "last player standing" crown term). Listing such a scalar in ACTIVE_ licenses CMA-ES to tune a win detector into noise.

When in doubt, ship in UNTUNED_ with a reason. Promotion is a one-line move backed by an ai-gate report; demotion after the optimizer has already smeared the value is not.

---

8. Plan/curves placement order (critical)

tempo_class_for in plan/curves.rs returns the FIRST matching TempoClass. Order matters because hybrid decks should resolve to a sensible class.

Established order (DON'T break this):

1. Ramp first (landfall.commitment > 0.5 OR mana_ramp.commitment > 0.5) → Ramp. Ramp+anything reads as Ramp.
2. Tribal (tribal.commitment > AGGRO_TEMPO_FLOOR) → Aggro. Tribal+aggro reads as Aggro (tribal wins position).
3. Aggro pressure (aggro_pressure.commitment >= AGGRO_TEMPO_FLOOR) → Aggro.
4. Tokens wide (tokens_wide.commitment >= TOKENS_WIDE_TEMPO_FLOOR) → Aggro.
5. Control (control.commitment > 0.55 AND control.reactive_tempo > 0.35) → Control.
6. Aristocrats (aristocrats.commitment > 0.5) → Midrange.
7. Plus-one counters (plus_one_counters.commitment > 0.5) → Midrange.
8. Default to archetype.

When inserting a new branch: ramp-style features go FIRST (they dominate), aggro/midrange/control go in the middle, "fallback" archetypes last. Use >= for consistency; the existing > in some branches is being migrated.

---

9. Recurring review traps (avoid these)

These are real bugs caught in the 9-feature batch's reviews. Each cost a follow-up commit.

Trap	Caught in	Fix pattern
CR drift — citing a rule number from memory that's actually a different rule	Counters review caught CR 122.1d (Stun, not +1/+1) and CR 701.27 (Transform, not Proliferate). Aggro caught CR 702.2 (Deathtouch, used for keyword field) and CR 800.4 (multiplayer, used for min-life helper).	Grep docs/MagicCompRules.txt for every CR before writing. Don't annotate plumbing.
Per-copy name push — payoff_names.push(name) inside for _ in 0..entry.count loop inflates the list to playset size	Tokens review caught (4× pushes per Bitterblossom).	Hoist the push outside the loop. Counts use entry.count; identity-lookup lists use one push per unique face.
Combo card double-push — if is_X { names.push() } if is_Y { names.push() } pushes the same name twice when both flags fire	Tokens review caught (Glorious Anthem + Overrun-shape).	Use if is_X || is_Y { names.push() } once per face.
Opponent-scoped trigger counted as your payoff — failing to reject valid_target = ControllerRef::Opponent makes "punisher" cards register as your payoffs	Counters review caught.	Mirror aristocrats::typed_filter_is_creature_you_control's if matches!(typed.controller, Some(ControllerRef::Opponent)) { return false; }.
Stack inspection only walks head effect — state.stack.iter().for_each(|e| check(&resolved.effect)) misses counter effects in sub_ability	Counters review caught.	Use super::context::collect_ability_effects(resolved) to walk the full ResolvedAbility chain.
Doubler check on activated source — passive replacements (Hardened Scales, Doubling Season) live on permanents, not on activated abilities	Counters review caught the dead branch.	Scan state.battlefield for permanents with replacement_definitions matching the predicate.
is_X wrapper unused — clippy flags wrappers without callers	Spellslinger reviewer noted; agent removed in implementation.

---

10. Cross-feature interaction patterns

When your feature should influence an existing policy, prefer surgical amplification over a new policy:

// Inside an existing policy's verdict() or score():
let amp = 1.0 + (features.<your_feat>.commitment as f64).clamp(0.0, 1.0) * 0.5;
penalty *= amp;

Examples shipped:

• tokens_wide amplifies BoardWipeTelegraphPolicy (wide-board decks fear sweepers more).
• mana_ramp::is_mana_dork_parts is consumed by HoldManaUpForInteractionPolicy (mana sources count untapped Sol Ring shapes).
• landfall::ability_searches_library_for_land is consumed by aristocrats (anti-fetchland gate prevents fetchlands counting as outlets).

When a helper needs to be shared across two features, promote it to pub(crate) in the owning module with a doc note ("Shared with X for Y"), don't recreate it.

---

11. Test plan template

Every feature MUST have:

Feature tests:

• empty_deck_produces_defaults
• vanilla_creature_not_registered
• detects_<each_axis> (one per detection axis)
• <exclusion>_does_not_count (negative cases — e.g., treasure_token_does_not_count, reach_does_not_count_as_evasion, loyalty_counter_not_counted)
• opponent_scope_<X>_ignored (controller rejection where applicable)
• <calibration_anchor_name>_hits_floor (positive calibration: assert!(commitment > 0.85))
• <negative_archetype>_below_floor (negative calibration)
• commitment_clamps_to_one (overflow saturation)
• <identity_list>_dedup (one push per unique face — see Section 9 trap)

Tactical policy tests:

• activation_opts_out_below_floor
• activation_opts_in_above_floor
• One test per verdict branch (positive scoring case)
• One test per verdict branch (negative/penalty case)
• non_<X>_action_returns_zero (<X>_na reason kind)

Mulligan policy tests:

• opts_out_when_commitment_low
• ideal_hand_<descriptors> → positive delta with <feat>_keepable_ideal reason
• Each verdict tier (workable, penalty, default)
• A turn_order / mulligans_taken test, or an explicit // input-unused: <reason> marker in the implementation.

Measurement evidence:

• New-policy PRs attach a rtk cargo ai-gate paired-seed report. No flips is acceptable for narrow policies, but the run must exist.
• If the policy calls any board-wide or affordability engine function (max_x_value, find_legal_targets, mana-availability / SimulationFilter sweeps), ALSO attach cargo ai-perf-gate (or a wall-clock A/B on a large late-game board). ai-gate measures win-rate only and is structurally blind to per-decision latency; ai-perf-gate is blind to any call that bumps no PerfCounterSnapshot field. See the Performance section.

---

12. Implementation order (each step compiles before the next)

1. Read landfall.rs + the most-similar reference feature end to end.
2. Verify each CR number you'll cite by grepping docs/MagicCompRules.txt.
3. STEP 0 (if needed): refactor an existing predicate from another feature to pub(crate) parts shape (e.g., control::is_card_draw_parts was promoted for spellslinger). Run the existing tests; behavior must be preserved.
4. Optionally start from scripts/new-ai-policy.sh <feat>; it writes the three scaffold files and prints the wiring checklist.
5. Create features/<feat>.rs: struct + thresholds + detect() + parts predicates + tests.
6. Wire features/mod.rs.
7. Create policies/<feat>.rs (tactical) + tests. Add PolicyId variant + Box::new() registration.
8. Create policies/mulligan/<feat>_keepables.rs + tests. Add PolicyId variant + register in MulliganRegistry::default.
9. (Optional) Add plan/curves.rs branches with tests.
10. (Optional) Cross-feature amplification (Section 10).
11. Run cargo fmt --all, then use Tilt verification: ./scripts/tilt-wait.sh clippy test-ai. If Tilt is down, fall back to the direct phase-ai clippy/test commands.
12. Run rtk cargo ai-gate for new policy work.
13. Commit with explicit git add path1 path2 .... NEVER git add -A.
14. Spawn an opus advisory review (see git log --grep="address review" for prior fix-up commit messages as templates). Address MUST FIX in a follow-up commit; CONSIDER items at your discretion.

---

13. Sample input contract (when invoking this skill)

When calling this skill or asking an agent to implement a new feature, provide:

1. Feature name (<feat>, snake_case).
2. Detection axes — one bullet per axis with the AST type to inspect (use the table in Section 3).
3. Calibration anchor — a real archetype's component counts so the formula can be sanity-checked.
4. Policies in scope — tactical only, or also mulligan, and any cross-feature amplification.
5. Boundary discussion — overlap with existing features (e.g., "tokens_wide overlaps with aristocrats on token generators; both axes are independent — a deck can score high in both").

The skill provides everything else (struct shape, formula choice, predicate signatures, CR table, wiring, tests, traps).