feat(analysis): combo-graph effect/trigger breadth + life-symmetry cost (Engine B, PR-4b)

[lgray]

🤖 AI text below 🤖

Combo-detector series

Part of the staged, offline-first infinite-combo / loop detector. Each PR links its predecessor so humans can follow the implementation trail end to end.

Pos	PR	Delivers
PR-0	#4092	ResourceVector + modulo-resource loop equality (additive, no behavior change).
PR-1	#4097	Analysis sim harness around GameRunner::act feeding ResourceVector.
PR-2	#4119	Net-progress detect_loop → LoopCertificate + loop-certificate corpus harness.
PR-3	#4480	Live mandatory-loop winner shortcut for drain-cascade combos (CR 704.5a).
PR-4a	#4493	Engine B static ability-graph extractor: graph scaffold + 5 effect families + SCC + candidate coverability.
PR-4b	#4534 (this PR)	Engine B effect/trigger breadth + life-symmetry cost.
PR-5	#4547	cargo combo-verify CLI over the 53-row corpus.

Predecessor: PR-4a — #4493 — #4493 (the static ability-graph extractor this PR broadens — more effect/trigger families + the life-symmetry cost half).

---

Engine B, PR-4b — combo-graph effect/trigger breadth + life-symmetry cost

Builds on PR-4a (#4493, the static ability-graph combo-candidate extractor) by broadening effect/trigger projection so more loop classes are detected, plus the cost half of the life-symmetry pair.

Effect projections

Draw (cards-drawn axis), Mill (negative opponent library), GainLife / LoseLife (sign-aware, per-player), token creation (tokens + ETB), Sacrifice / Destroy / ChangeZone (sac / LTB / death-trigger disambiguation by filter and zone transition), and extra-turn / extra-combat-phase axes.

Trigger axes

ChangesZone disambiguation (ETB vs death vs LTB) and event triggers requiring their axes.

Cost

PayLife now folds as the cost half of the R3-LIFE-SYMMETRY pair (with GainLife/LoseLife) — its own fold_cost arm (CR 119.4), so a gain-and-pay-life loop isn't vetoed as net-negative. Composite stays AND-fold; OneOf stays the disjunctive fold_one_of from PR-4a (CR 118.12a).

Together these detect e.g. dies-token aristocrats loops and lifegain-feedback loops as candidates.

Drift gates / typing

fold_cost, project_mana_production, and trigger_axis remain exhaustive no-wildcard drift gates; model completeness uses the typed ModelCompleteness enum (PR-4a). This PR is rebased onto current main (which includes the merged PR-4a review fixes — fold_one_of, ModelCompleteness, net-sign axis_key_to_resource); the reconciliation merged PR-4b's breadth with those structures and migrated the one test that referenced the removed any_unmodeled field.

Tests

17 PR-4b breadth/fixture tests + the 4 PR-4a-review tests; ability_graph module 40/40. cargo test -p engine 14007 passed / 0 failed; cargo clippy --workspace --exclude phase-tauri --all-targets --features engine/proptest -- -D warnings = 0. All CR annotations grep-verified.

[gemini-code-assist]

Code Review

This pull request implements the PR-4b phase of the static analysis engine, adding support for projecting and modeling various game effects and trigger modes (such as Draw, Mill, Life, Tokens, Zone-Changes, Sacrifice, and Extra Turns/Phases) onto resource event axes. It also updates trigger axis mapping to exhaustively handle these new event triggers and introduces comprehensive unit tests for aristocrats and lifegain feedback loops. The review feedback suggests two improvements: enhancing type_filters_exclude_creature to support disjunctive non-creature constraints (e.g., Non(AnyOf(...))) to avoid spurious Death edges, and updating the Effect::Conjure projection to handle multi-card and variable-count conjuring rather than hardcoding a single ETB event.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[matthewevans]

Reviewed current head 2b8699723a066db142c20f3ed6e389411fb75620 in worktree /tmp/forge-pr-4534, including the unresolved review threads and the only changed file (crates/engine/src/analysis/ability_graph.rs). I agree this needs another pass before it is ready.

[MED] type_filters_exclude_creature only recognizes direct Non(Creature) and misses composed exclusions. Evidence: crates/engine/src/analysis/ability_graph.rs:397-400 only checks for TypeFilter::Non(inner) where inner is exactly Creature, while the engine type model supports TypeFilter::Non(Box<TypeFilter>) and TypeFilter::AnyOf(Vec<TypeFilter>) (crates/engine/src/types/ability.rs:2434-2458). A filter like Non(AnyOf([Creature, ...])) also provably cannot match a creature, but this helper will still project Sacrifice/Destroy as producing Death. Please make this reasoning recursive over the existing TypeFilter composition shape, preserving the conservative behavior for positive non-creature card types.

[MED] Effect::Conjure projection collapses all battlefield conjures to one fixed ETB. Evidence: crates/engine/src/analysis/ability_graph.rs:771-773 ignores the cards field and always calls b.add_etb(1, AxisMagnitude::Fixed(1)), but Effect::Conjure carries cards: Vec<ConjureCard> and each ConjureCard has a count: QuantityExpr (crates/engine/src/types/ability.rs:7509-7521, 10462-10465). Multi-card or counted conjure effects will be under-projected in the combo graph. Please aggregate the fixed counts where possible and mark variable/unknown quantities conservatively, similar to the existing count_seed path used by token creation.

[lgray]

🤖 AI text below 🤖

Both addressed in e5511a13a (rebased onto current main).

[MED] type_filters_exclude_creature — now recursive over Non/AnyOf. Replaced the direct Non(Creature)-only check with two mutually-recursive predicates over the full TypeFilter shape (exhaustive, no-wildcard):

• type_filter_excludes_creature(tf) — provably matches no creature: Non(inner) iff inner matches every creature; AnyOf(types) iff every branch excludes; positive types/subtypes stay conservative false (CR 205.2b — a creature can also be a land/artifact/etc.).
• type_filter_matches_all_creatures(tf) — the dual used for Non: Creature/Any → true; Non(inner) iff inner excludes; AnyOf iff some branch matches all.

So Non(AnyOf([Creature, Land])) ("neither a creature nor a land") is now correctly recognized as excluding creatures, while Non(AnyOf([Land, Artifact])) is not (a creature can satisfy it) — and it composes to arbitrary nesting. Added type_filter_excludes_creature_handles_composed_negation, which is discriminating (the pre-fix direct match returns false for Non(AnyOf([Creature, Land]))).

[MED] Effect::Conjure projection — now per-ConjureCard count. Replaced the single fixed ETB with a loop over cards, seeding the ETB axis from each ConjureCard.count via count_seed (mirroring the token-creation path) — multi-card and counted conjures aggregate their fixed counts, and variable/X counts mark the axis Unbounded. Non-battlefield destinations stay Unmodeled (CR 603.6a).

cargo test -p engine 14008/0; cargo clippy --workspace … --features engine/proptest -- -D warnings 0.

[matthewevans]

Reviewed current head e5511a13af2bc41628d0f00fd55f0ab5e1e384df. The two issues from my prior review look addressed, but one sibling gap remains.

[MED] LifeLostAll should consume the Life axis too. Evidence: crates/engine/src/analysis/ability_graph.rs:1047 maps LifeLost/LifeChanged/PayLife to AxisKey::Life, but crates/engine/src/analysis/ability_graph.rs:1099 leaves LifeLostAll in the inert None bucket; runtime treats LifeLostAll as the same life-loss event class in crates/engine/src/game/trigger_matchers.rs:63 and indexes it with life-change triggers in crates/engine/src/game/trigger_index.rs:259. Why it matters: Engine B will miss candidates whose consumer is the all/batched life-loss trigger form even though this PR now models the corresponding life-loss producers. Suggested fix: include TriggerMode::LifeLostAll in the Life branch and add a discriminating trigger_axis or candidate-cycle test for a LifeLostAll trigger fed by a life-loss producer.

[lgray]

🤖 AI text below 🤖

Good catch on the sibling gap — fixed in the latest push (rebased onto current main).

[MED] LifeLostAll now consumes the Life axis. Moved TriggerMode::LifeLostAll from the inert None bucket into the Life branch of trigger_axis, alongside LifeLost/LifeChanged/PayLife. This matches the runtime, which routes LifeLostAll through the same match_life_lost matcher as LifeLost (trigger_matchers.rs) and indexes it with the life-change triggers (trigger_index.rs) — so Engine B now connects the batched/all life-loss consumer to the life-loss producers this PR models.

Added life_lost_all_trigger_consumes_life_axis, which is discriminating (pre-fix LifeLostAll returned None from trigger_axis).

cargo test -p engine 14011/0; cargo clippy -D warnings 0.

[matthewevans]

Approved current head 822c061e11bc6eda6cd9065c34f2b99fa1c5260a after re-review. The prior LifeLostAll trigger-axis blocker is fixed; the earlier recursive non-creature filter and per-card Conjure ETB findings remain addressed; exact-head CI is green.

Evidence checked: one-file backend diff in crates/engine/src/analysis/ability_graph.rs; unresolved review-thread gate is clear; no workflow/instruction/security-sensitive paths; no duplicate open PR surfaced by title/issue scan.