feat(parser): lift "[once] for each ⟨player-set⟩" onto fieldless Investigate

[lgray]

🤖 AI text below 🤖

Summary

Generalize the fieldless-Effect::Investigate "for each ⟨set⟩" lift from PlayerCount-only to the parameterized member-count class (PlayerCount | ObjectCount), and add the FilterProp::Goaded building block it needs. Addresses the #6110 review (the lift dropped object-ranged Investigate to a bare 1-Clue, rules-wrong CR 701.16a/608.2c).

Net effect: 7 real parse (AST) changes, all strict improvements, ZERO regressions (confound-free regen-both card-data.json diff at the merge-base 931c2dc2f). The repeat_for coverage projection is scoped to the lift class (see below), so the coverage-parse-diff sticky reports 8 cards / 10 signatures — those 7 plus one pre-existing, AST-identical card (Ethereal Investigator) whose repeat_for merely surfaces in the scoped signature.

• 3 intended repeat_for lifts: Teysa, Opulent Oligarch (PlayerCount, opponents who lost life) · Wojek Investigator (PlayerCount, comparative hand-size) · Serene Sleuth (ObjectCount, goaded creatures you control).
• 4 collateral goaded-filter improvements (the general FilterProp::Goaded correctly parsing "goaded creature" where it was previously Unknown/degraded — inherent to a general building block): Bothersome Quasit (CantBlock static) · Puppet Master, String Puller · The Rani · Vengeful Ancestor (goaded-subject triggers).
• Deliberately NOT lifted: Tamiyo Meets the Story Circle's "investigate twice for each card discarded" (a contextual amount-ref) — lifting it would silently drop the "twice" multiplier. A fail-closed matches!(PlayerCount | ObjectCount) eligibility gate excludes it, locked by a revert-probed negative-boundary test. Byte-identical to base.

How this addresses the #6110 review

• Blocker 1 (generalize the lift to retain every supported QuantityRef incl. ObjectCount + a runtime discriminator for ≥1 object-set card) → lower.rs gate widened PlayerCount-only → matches!(PlayerCount | ObjectCount); FilterProp::Goaded added so "goaded creature you control" parses to ObjectCount; the :404 enshrining test inverted to require the object-set lift. Runtime discriminator: serene_sleuth_runtime_makes_one_clue_per_goaded_creature (3 goaded → 3 Clues vs Goaded-blind 5 vs bare 1, revert-probed). Serene Sleuth is fixed and proven.
• Blocker 2 + the coverage blast-radius follow-up → the repeat_for projection in coverage.rs ability_details is scoped to the lift class — it surfaces the repeat_for signature key only for a fieldless Effect::Investigate whose repeat_for is a member-count QuantityRef (PlayerCount/ObjectCount), exactly the eligibility set of for_each_repeatable_repeat_for. It does not touch the rest of the repeat_for surface (CopySpell/Token/Proliferate/SearchLibrary, or pre-existing Fixed/Variable/tracked-set Investigate forms). Measured with the real coverage-parse-diff tool against the merge-base (931c2dc2f, no-projection baseline): 8 cards / 10 signatures (down from the earlier global projection's 250), confound-free (0 oracle-text changes, 0 added/removed). That is 7 real parse changes (the 3 lifts Teysa/Wojek/Serene + the 4 collateral goaded-filter cards Bothersome Quasit, Puppet Master, The Rani, Vengeful Ancestor) plus 1 projection-only, AST-identical card — Ethereal Investigator, a pre-existing Investigate + PlayerCount repeat_for whose value simply becomes visible in the scoped signature (byte-identical card-data.json; structurally indistinguishable from the lifts, so no non-hacky scope excludes it). A revert-probed investigate_signature_exposes_repeat_for regression (2 positive member-count cases project; 4 negatives — None, Fixed, non-member-count Ref, non-Investigate effect — do not) locks the scope. Broadening repeat_for coverage to the whole surface is a deliberate global coverage-schema migration, deferred out of this focused feature.
• Goaded event-snapshot certificate (matthewevans's 4th review, [HIGH]) — FIXED (option b). classify_prop in types/events.rs had FilterProp::Goaded in the Supported group, asserting the reach gate could answer a goaded event-subject filter from an EventObjectSnapshot. But the snapshot carries no goaded field and the runtime zone-change-record matcher already fail-closes it (game/filter.rs ~:5049, CR 701.15b/c) — an inconsistent, fail-open certificate. Reclassified FilterProp::Goaded → Unsupported, aligning the certificate with the snapshot's actual fields and the runtime's existing fail-closed behavior, so a future goaded event-subject filter fails the reach gate loudly rather than silently reading an ungoaded snapshot. Pure contract+test change with zero behavior/coverage delta: the reach gate is test-only latent scaffolding (classify_prop/classify_filter_shape have no production caller today), coverage is supplied independently by coverage.rs, and the live goaded predicate (filter.rs FilterProp::Goaded => !obj.goaded_by.is_empty()) is untouched. Measured: 4-card and full-DB card-data.json byte-identical before/after, plus a revert-probed goaded_subject_filter_is_unsupported test (returning Goaded to Supported measurably flips the assertion). Option-(a) snapshot-plumbing path is deferred (see Deferred follow-ups).
• Sophina, Spearsage Deserter — disclosed follow-up, NOT fixed here. It remains a bare Investigate. Its "nontoken attacking creature" filter is blocked by a distinct, pre-existing parse_type_phrase leading-adjective order-dependence (the nontoken-leads-attacking ordering returns None before the lift seam is reached), orthogonal to the gate fixed above. Deferred as a strict-failure tag + a flipping tripwire test (the Sophina branch of the inverted :404 regression), because the order-free many0 fix is the hottest shared parser path. The blocker is answered at the seam level (ObjectCount retained, proven by Serene Sleuth); Sophina is a scoped follow-up.

Implementation method (required)

Method: /engine-implementer

CR references

• CR 608.2c — resolve the effect's instructions; the repeat_for loop reruns the fieldless Investigate once per iteration.
• CR 701.16a — "Investigate" = create a Clue token (CR 111.10f).
• CR 701.15b/c — goad designation (FilterProp::Goaded reads GameObject.goaded_by; multi-player set).
• CR 508.2a + CR 603.2 — an attacks-triggered ability's trigger event is the attacking creature; its valid_card filter is evaluated against the live attacker.
• CR 109.5 — "you"/"than you" binds the controller (Wojek's comparative operand).
• CR 402.1 — cards in hand (HandSize).

Verification

• Required checks ran clean, or the exact CI-owned alternative is stated below.

• Gate A output below is for the current committed head.

• Final review-impl below is clean for the current committed head.

• Both anchors cite existing analogous code at the same seam.

• scripts/check-parser-combinators.sh upstream/main — Gate G + Gate A PASS (head 29e9ad9 base 931c2dc)

• cargo fmt --all -- --check — clean

• cargo clippy --workspace --all-targets -- -D warnings — clean (0 warnings; post-rebase drift-verify) + cargo clippy -p engine --test integration -D warnings clean for the new test

• cargo test -p engine --lib — 40 passed (goaded / for_each_repeatable / dq_d_player_set_lift / investigate_signature + upstream goaded-combat)

• cargo test -p engine --test integration — feature subset passed (teysa/wojek runtime, serene runtime 3-way, object_for_each lift+Sophina-deferred guard, vengeful goaded-attack pair)

• Regen-both card-data.json diff (base 931c2dc2f = merge-base, feature head; same MAIN corpus symlinked into both worktrees, full per-card face diff): 35,396 cards both sides, no only-in-X ⇒ confound-free. 7 real AST changes, all strict improvements, 0 regressions. Byte-identity confirmed on Tamiyo Meets the Story Circle, Confront the Unknown, Declaration in Stone, Tranquil Frillback.

• coverage-parse-diff tool (real binary, base 931c2dc2f no-projection snapshot vs the scoped-projection head; base verified: 0 repeat_for keys, 15 Investigate+repeat_for with the 3 lifts absent = no-lift): 8 cards / 10 signatures, confound-free (oracle_changed=0, added=0, removed=0). = the 7 AST cards + Ethereal Investigator (projection-only, AST-identical). Down from the earlier global projection's 250.

Runtime discrimination (apply()-drives, revert-probed)

• Serene Sleuth (ObjectCount count context): serene_sleuth_runtime_makes_one_clue_per_goaded_creature — 3 goaded + ungoaded Sleuth + 1 plain → 3 Clues (vs Goaded-blind 5, vs bare 1); reverting the Goaded eval flips it.
• Vengeful Ancestor (trigger-subject context — a distinct wire): vengeful_ancestor_goaded_attack_trigger.rs — goaded attacker → controller loses 1 (trigger fires); ungoaded attacker → no loss (does not fire); legs differ only in goaded state. Production-line revert-probe (filter.rs FilterProp::Goaded => false) flips the goaded leg. Proves the trigger valid_card Goaded filter resolves against the live attacker's goaded_by, not the fieldless EventObjectSnapshot — so the collateral cards are genuinely supported, not false-supported.
• Teysa / Wojek (PlayerCount): teysa_runtime_*, wojek_runtime_*.

The repeat_for driver drive_repeat_for_outermost (crates/engine/src/game/effects/mod.rs) resolves the quantity generically with zero per-filter branching.

Gate A

Gate A PASS head=29e9ad967 base=931c2dc2f

Anchored on

• crates/engine/src/parser/oracle_effect/lower.rs — for_each_repeatable_repeat_for (renamed/widened from for_each_player_set_repeat_for) shares the split_for_each_suffix primitive with the pre-existing strip_for_each_repeat_suffix lift at the same seam.
• crates/engine/src/parser/oracle_effect/mod.rs — the existing repeat_for: Some(QuantityExpr::Ref{..}) chunk-seam lift the widened else if let Some(..) branch follows.
• crates/engine/src/game/filter.rs / types/ability.rs — FilterProp::Goaded mirrors FilterProp::Renowned across all 15 registration sites (candidate-local read of goaded_by).

Final review-impl

Final review-impl PASS head=29e9ad967 (independent /review-impl, zero blocking findings; the collateral live-eval risk was subsequently measured and cleared by the Vengeful Ancestor runtime pair).

Claimed parse impact

coverage-parse-diff reports 8 cards / 10 signatures (confound-free vs merge-base 931c2dc2f):

7 real parse (AST) changes:

• Teysa, Opulent Oligarch
• Wojek Investigator
• Serene Sleuth
• Bothersome Quasit
• Puppet Master, String Puller
• The Rani
• Vengeful Ancestor

1 projection-only, AST-identical (disclosed, not a regression):

• Ethereal Investigator — pre-existing Investigate + PlayerCount repeat_for; byte-identical card-data.json, surfaces only because the scoped coverage signature now shows its repeat_for. Structurally indistinguishable from the lifts, so no non-hacky scope excludes it.

Deferred follow-ups

Recorded in-code as a strict-failure tag with a flipping tripwire test (no silent drop):

• parse_type_phrase leading-adjective order-dependence (Sophina, Spearsage Deserter — "nontoken attacking creature"). The leading-adjective stripping is a fixed positional cascade (combat-status → … → nontoken); a phrase whose adjectives cross that order is only partly stripped, so parse_for_each_clause returns None and the per-attacker Investigate stays honestly bare (no false lift). Fix: collapse the cascade into an order-free many0-style property loop — the hottest shared parser path (high CI-coverage-regression blast radius), hence deferred. Tripwire: the Sophina branch of object_for_each_investigate_is_lifted asserts the current bare-Investigate state and flips to fail when the gap closes.
• EventObjectSnapshot goaded-projection certificate — resolved in this PR (option b, fail-closed reclassify; see the review-addressing section above). The remaining deferred (option a) is the fuller fix: snapshot goad onto EventObjectSnapshot + ZoneChangeRecord (mirroring is_suspected), wire the capture sites, read it at both runtime matchers, and reclassify FilterProp::Goaded back to Supported. Deferred because no card today prints a goad predicate on an event-subject filter under last-known-information — the reach gate is currently unreachable and goaded_by clears on zone change, so the fail-closed classification is correct and complete for every current card. FilterProp::Renowned is NOT the same shape (correcting an earlier note): EventObjectSnapshot already carries is_renowned, so Renowned=Supported is consistent at this reach gate; Renowned's only gap is on the separate zone-change-record path — a distinct, pre-existing concern, out of scope for this PR.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[lgray]

🤖 AI text below 🤖

@matthewevans ready for review + enqueue when you have a moment. This closes the same missing-wire class you flagged on #6108 (Bumi): a parser-only lift of a trailing "[once] for each ⟨player-set⟩" clause onto a fieldless Effect::Investigate, plus the end-to-end regression you would want — two apply()-drives (teysa_runtime_makes_one_clue_per_opponent_who_lost_life, teysa_runtime_no_clue_when_no_opponent_lost_life) with revert-probe evidence: disabling the wire flips the Clue count 0→1 and trips the reach-guard.

Gates green on head 67da88d (rebased on current main 7afcff9): clippy --all-targets -D warnings, integration 7/7, lib 3/3, Gate A, all 6 CRs resolve, coverage-diff = exactly {Teysa, Wojek} (+2, 0 regressions). I am pull-only so I cannot self-enqueue — over to you for the merge queue. Thanks!

[matthewevans]

Maintainer update: the implementation re-review on the current head is clean. Approval remains held until GitHub publishes the current-head coverage-parse-diff artifact and the required Rust lint/test checks finish successfully. No action is needed from you right now; I’ll resume once that evidence lands.

[github-actions]

Parse changes introduced by this PR · 8 card(s), 10 signature(s) (baseline: main c7225018447d)

🟢 Added (2 signatures)

• 2 cards · ➕ trigger/DamageDone · added: DamageDone (active in=battlefield, damage kind=combat only, valid source=goaded creature, valid target=opponent)
  • Affected (first 3): Puppet Master, String Puller, The Rani
• 1 card · ➕ trigger/Attacks · added: Attacks (active in=battlefield, watches=goaded creature)
  • Affected (first 3): Vengeful Ancestor

🔴 Removed (3 signatures)

• 1 card · ➖ trigger/Whenever a goaded creature attacks · removed: Whenever a goaded creature attacks (active in=battlefield)
  • Affected (first 3): Vengeful Ancestor
• 1 card · ➖ trigger/Whenever a goaded creature deals combat damage to one of your opponents · removed: Whenever a goaded creature deals combat damage to one of your opponents (active in=battlefield)
  • Affected (first 3): The Rani
• 1 card · ➖ trigger/Whenever one or more goaded creatures deal combat damage to one of your opponen… · removed: Whenever one or more goaded creatures deal combat damage to one of your opponents (active in=battlefield)
  • Affected (first 3): Puppet Master, String Puller

🟡 Modified fields (5 signatures)

• 1 card · 🔄 static/CantBlock · changed field affects: self → goaded opponent controls creature
  • Affected (first 3): Bothersome Quasit
• 1 card · 🔄 ability/Investigate · changed field repeat_for: ∅ → # of each opponent
  • Affected (first 3): Ethereal Investigator
• 1 card · 🔄 ability/Investigate · changed field repeat_for: ∅ → # of each opponent who lost life this turn
  • Affected (first 3): Teysa, Opulent Oligarch
• 1 card · 🔄 ability/Investigate · changed field repeat_for: ∅ → # of each opponent whose HandSize { player: ScopedPlayer } GT Ref { qty: HandSize { player: Controller } }
  • Affected (first 3): Wojek Investigator
• 1 card · 🔄 ability/Investigate · changed field repeat_for: ∅ → # of goaded you control creature
  • Affected (first 3): Serene Sleuth

2 card(s) had Oracle-text changes (errata/reprint) — excluded as non-parser.

[matthewevans]

Request changes — the player-set lift is at the right chain-lowering seam, but it deliberately leaves the same fieldless investigate for each … grammar rules-incorrect for object sets.

🔴 Blocker

• crates/engine/src/parser/oracle_effect/lower.rs:3581-3586 accepts only QuantityRef::PlayerCount; the new crates/engine/tests/integration/teysa_wojek_investigate_per_opponent.rs:404 regression expressly permits object-set investigate for each to remain without a repeat count. The verified Oracle text for Serene Sleuth is “investigate for each goaded creature you control,” and Sophina, Spearsage Deserter says “investigate once for each nontoken attacking creature.” repeat_for already models an integer count, so neither needs per-object binding: both must create one Clue per matching creature, not silently resolve as one Investigate. CR 701.16a defines investigate as creating a Clue; CR 608.2c requires following the full instruction. Generalize the fieldless-effect suffix lift to retain every supported QuantityRef (including ObjectCount) and add a runtime discriminator for at least one object-set card.

🟡 Non-blocking

• The current-head <!-- coverage-parse-diff --> artifact says “No card-parse changes detected,” whereas the PR comment claims exactly {Teysa, Wojek} (+2). This prevents the manual Quality Gate from passing or a quality label; it is not evidence that the behavior change is absent. Reconcile the claimed parse-impact statement after the blocker is fixed.

✅ Clean

• The change is in parse_effect_chain_ir, the appropriate repeat attachment seam, and the Teysa/Wojek tests exercise real parse and scenario-resolution paths.

Recommendation: extend the lift across the complete supported fieldless-Investigate for each class (player and object counts), then re-request review with the regenerated parse-diff evidence.

[matthewevans]

The current-head parse-diff sticky says “No card-parse changes detected,” but that is a coverage false-green rather than evidence that this change is inert.

This PR attaches repeat_for to otherwise fieldless Effect::Investigate for the Teysa/Wojek player sets (oracle_effect/mod.rs), so the parse tree materially changes. The coverage parse tree cannot report it: build_ability_item merges effect_details and ability_details, but ability_details does not project AbilityDefinition::repeat_for (crates/engine/src/game/coverage.rs:3592), and Effect::Investigate deliberately has no effect detail (:3521). The separate feature extractor does recognize repeat_for (:7050), but that does not feed the parse-diff sticky.

Please project repeat_for (including its quantity) into ability parse details and add a regression proving this Teysa/Wojek class produces a parse-diff signature. Then the CI artifact can serve as the claimed card-level evidence.

[lgray]

🤖 AI text below 🤖

Addressed the two requested changes.

Blocker 1 — generalize the fieldless-Investigate "for each" lift beyond PlayerCount. lower.rs now retains the full member-count class matches!(PlayerCount | ObjectCount), and the :404 regression was inverted to require the object-set lift. Serene Sleuth ("investigate for each goaded creature you control") lifts to a runtime-resolved ObjectCount via a new FilterProp::Goaded building block (mirrors FilterProp::Renowned), with the runtime discriminator requested: serene_sleuth_runtime_makes_one_clue_per_goaded_creature — 3 goaded creatures → 3 Clues (vs 5 if the Goaded filter were ignored, vs 1 for a bare Investigate), revert-probed on the production FilterProp::Goaded eval. A fail-closed gate keeps contextual amount-refs out: Tamiyo Meets the Story Circle's "investigate twice for each card discarded" stays byte-identical (lifting it would silently drop the "twice"), locked by a negative-boundary test.

Blocker 2 — project repeat_for into the coverage parse tree. coverage.rs ability_details now pushes ("repeat_for", fmt_quantity(rf)) (None → byte-identical, no spurious diff), with an investigate_signature_exposes_repeat_for regression (Some → present / None → absent, revert-probed). This also resolves the non-blocking parse-diff false-green: a confound-free regen-both (base = feature merge-base) now reports real card-parse changes — 7 cards, all strict improvements, zero regressions (the 3 intended lifts plus 4 collateral goaded-filter cards whose previously-Unknown "goaded creature" filters now parse, with a revert-probed runtime drive proving the trigger-subject Goaded eval resolves against the live object). The claimed parse-impact is reconciled to those 7.

Deferred — Sophina, Spearsage Deserter. It remains a bare Investigate (one Clue). Its "nontoken attacking creature" filter is blocked by a separate, pre-existing parse_type_phrase leading-adjective order-dependence — the nontoken-leads-attacking ordering makes the type parse return None before the lift seam is ever reached, so it is orthogonal to the gate fixed here. It is deferred as a documented strict-failure tag with a flipping tripwire test (the Sophina branch of the inverted :404 regression), rather than silently, because collapsing that fixed adjective cascade into an order-free loop is the hottest shared parser path and warrants its own change.

[matthewevans]

Maintainer update: I have received the new head and its response to the prior requested changes. I am holding the implementation review until GitHub publishes the current-head coverage-parse-diff artifact and the in-progress Rust checks complete; no further contributor action is requested at this point.

[matthewevans]

Request changes — the object-count runtime fix may be sound, but the current-head parse-diff exposes a 250-card coverage blast radius that the claimed seven-card scope does not account for.

🔴 Blocker

[HIGH] repeat_for coverage projection changes the parse signature of the whole existing repeat-for surface. The current <!-- coverage-parse-diff --> reports 250 changed cards / 158 signatures, including many unrelated existing repeat_for effects (for example CopySpell, Token, Proliferate, SearchLibrary, and dozens of Investigate forms), while the PR describes exactly seven card changes. The new coverage.rs projection therefore makes this focused Investigate/Goaded feature alter card-level parse output across the global repeat-for population.

Why it matters: the advertised parse-impact evidence is no longer a bounded witness for the intended class, and downstream parse-diff consumers now see hundreds of unrelated changes. Suggested fix: restrict the coverage signature change to the intended fieldless-Investigate lift class, or explicitly split and independently validate a deliberate global coverage-schema migration with a complete compatibility/consumer audit and reconciled artifact.

✅ Clean

• The current artifact does show Serene Sleuth, Teysa, Wojek, and the goaded-filter cards as real changes; this finding concerns the additional 243 cards.

Recommendation: split or narrow the coverage projection, then regenerate the parse-diff and re-request review.

[lgray]

🤖 AI text below 🤖

Addressed the coverage parse-diff blast-radius finding.

The repeat_for coverage projection is now scoped to the fieldless-Investigate lift class — coverage.rs ability_details surfaces the repeat_for signature key only for an Effect::Investigate whose repeat_for is a member-count QuantityRef (PlayerCount/ObjectCount), i.e. exactly the eligibility set of the lift (for_each_repeatable_repeat_for). It no longer touches the rest of the repeat_for surface (CopySpell/Token/Proliferate/SearchLibrary, or the pre-existing Fixed/Variable/tracked-set Investigate forms).

Measured with the coverage-parse-diff tool against the merge-base (931c2dc2f, the no-projection baseline), the scoped projection reports 8 cards / 10 signatures (down from the global projection's 250), confound-free (0 oracle-text changes, 0 added/removed cards):

• 7 real parse changes — the 3 repeat_for lifts (Teysa, Wojek, Serene) + 4 collateral goaded-filter cards (Bothersome Quasit, Puppet Master, The Rani, Vengeful Ancestor).
• 1 projection-only, AST-identical — Ethereal Investigator, a pre-existing Investigate + PlayerCount repeat_for whose value now surfaces in the scoped signature; its parse is byte-identical to the merge-base (disclosed, not a regression), and it is structurally indistinguishable from the lifts, so no non-hacky scope excludes it.

A revert-probed investigate_signature_exposes_repeat_for regression (2 positive member-count cases project; 4 negatives — None, Fixed, non-member-count Ref, non-Investigate effect — do not) locks the scope. Broadening repeat_for coverage to the whole surface would be a deliberate global coverage-schema migration, deferred out of this focused feature.

[matthewevans]

**Request changes — the scoped coverage projection resolves the prior blast-radius finding, but FilterProp::Goaded is now declared snapshot-answerable without snapshotting the designation.

🔴 Blocker

[HIGH] Event/LKI snapshot filtering silently treats a goaded subject as ungoaded. Evidence: crates/engine/src/types/events.rs:301-344 has no goad field, while :562-574 classifies FilterProp::Goaded as Supported; the live predicate reads obj.goaded_by at crates/engine/src/game/filter.rs:4302-4307, and the LKI adapter constructs a ZoneChangeRecord without a goad designation at :1509-1555. Why it matters: any reachable past-event or LKI subject filter using goaded becomes false after the object leaves the battlefield, despite the structural gate representing it as supported. Suggested fix: carry a capture-time goaded fact through the snapshot/record evaluators (and their serialization/construction sites), or classify this predicate unsupported until that full path exists; add a departure/LKI discriminating regression.

Recommendation: request changes for the snapshot contract, then re-run the current-head review.

[lgray]

🤖 AI text below 🤖

Addressed the [HIGH] EventObjectSnapshot goaded-projection certificate finding (option b — fail-closed reclassify).

classify_prop in types/events.rs classified FilterProp::Goaded as Supported, asserting the event-subject reach gate could answer a goaded filter from an EventObjectSnapshot. It can't: the snapshot carries no goaded field, and the runtime zone-change-record matcher already fail-closes Goaded (game/filter.rs ~:5049, CR 701.15b/c). The certificate was fail-open/inconsistent.

Fix: reclassified FilterProp::Goaded → Unsupported, aligning the reach-gate certificate with the snapshot's actual fields and the runtime's existing fail-closed behavior. A future card that prints a goad predicate on an event-subject filter now fails the reach gate loudly (the designed "extend the snapshot + evaluator together" signal) instead of silently reading an ungoaded snapshot.

This is a pure contract + test change with zero behavior/coverage delta:

• The reach gate is test-only latent scaffolding — classify_prop/classify_filter_shape have no production caller today (referenced only inside events.rs, incl. a #[cfg(test)] helper).
• Coverage for goaded cards is supplied independently by coverage.rs (FilterProp::Goaded => parts.push("goaded")), not by classify_prop.
• The live goaded predicate (filter.rs FilterProp::Goaded => !obj.goaded_by.is_empty()) is untouched, so the PR's goaded triggers/statics (Vengeful Ancestor, Puppet Master, The Rani, Bothersome Quasit) resolve exactly as before.

Evidence:

• Revert-probed test goaded_subject_filter_is_unsupported — Typed{Creature,[Goaded]} classifies Unsupported. Measured revert-probe: returning Goaded to the Supported group flips the test to FAILED (assertion left == right failed, left: Supported, right: Unsupported); restoring → GREEN. Non-vacuous because TypeFilter::Creature classifies Supported, so combine_conjunction does not short-circuit the type axis — the flip is attributable to classify_prop(Goaded) alone.
• No-regression, byte-identical card-data.json — 4-card spot-check (the 4 goaded cards) and full-DB regen both cmp-identical before/after the change.
• cargo fmt / cargo clippy -p engine --lib --tests -D warnings clean.

Deferred follow-up (option a): snapshot goad onto EventObjectSnapshot + ZoneChangeRecord (mirroring is_suspected) and reclassify back to Supported — only needed once a card actually prints such a filter. FilterProp::Renowned is a separate, pre-existing concern (its snapshot field is_renowned already exists, so it is consistent at this gate; its only gap is on the distinct zone-change-record path), out of scope here.

[matthewevans]

Approved on the updated current head. The prior blocker is resolved by classifying goaded event-subject filters as snapshot-unsupported, so the existing runtime fail-closed behavior is reflected by the reach gate rather than silently certifying a missing LKI field. The maintainer branch update merged current main without a semantic conflict; merge queue will wait for the restarted checks.