[Bug]: AI legal-target check ignores Protection/HexproofFrom (removal fizzles)

[dearsishs]

Description

The AI's anti-self-harm policy and tactical gate decide whether an opponent creature is a legal target using a hand-rolled !has_keyword(Hexproof) && !has_keyword(Shroud) check. This ignores Protection(_), HexproofFrom(_), and the controller-relative nature of hexproof. The engine already exposes the CR-correct find_legal_targets, which the AI should use instead of duplicating targeting legality.

Steps to reproduce

1. Give an opponent a creature with Protection(white) (or HexproofFrom(white)).
2. Give the AI a white removal/bounce spell (e.g. Murder).
3. Let the AI evaluate casting it.

Expected behavior

The AI recognizes there is no legal target (the protected creature can't be targeted, CR 702.11/702.16/702.18) and does not score the spell as a productive cast.

Actual behavior

The AI treats the protected creature as targetable, casts the spell, and it fizzles on resolution (CR 608.2b) — a wasted card and lost tempo.

Impact

Removal/bounce mis-targeting is a visible bad-play signature; it also duplicates targeting legality in the AI in violation of "the engine owns all logic".

Component

• crates/phase-ai/src/policies/anti_self_harm.rs (has_targetable_opponent_creature, has_opponent_bounce_target, ETB-whiff detector)
• crates/phase-ai/src/tactical_gate.rs (is_redundant_creature_only_removal)
• crates/engine/src/game/targeting.rs (find_legal_targets — the CR-correct check to reuse)

[mike-theDude]

Additional reproduction from gameplay:

• Player controls Cryptothrall, which grants hexproof to other artifact creatures they control.
• The same player controls Ultron, Artificial Malevolence, an artifact creature, so Ultron has regular Hexproof from Cryptothrall.
• The AI opponent cast Dismantling Wave and was able to choose Ultron as the target artifact.

Expected: Dismantling Wave should not be able to target Ultron because hexproof prevents opponents from targeting that permanent.

This confirms the issue is not just Protection/HexproofFrom evaluation. The AI is also presenting or choosing targets that violate regular controller-relative Hexproof granted by a static ability.

[mike-theDude]

Implemented a fix for the AI Dismantling Wave / Cryptothrall / Ultron repro in commit 2fdcad18d.

The engine per-opponent fanout path now binds "that player controls" to the named opponent while still checking hexproof / hexproof-from / protection against the actual spell or ability controller. Added fanout regressions for regular hexproof, hexproof-from, ignore-hexproof bypass, resolution revalidation, and trailing sub-ability targets.

Verification: cargo fmt --all, git diff --check, cargo test -p engine fanout, cargo clippy --all-targets -- -D warnings.

[mike-theDude]

Reviewed against current main: the original AI legality scope was fixed by b2093b363, which routes removal, bounce, ETB, and tactical-gate checks through engine target legality with Protection and HexproofFrom regressions. The later Dismantling Wave reproduction from this thread is also covered on main by 3ed174201 and focused fanout/hexproof tests. The full reported scope now appears resolved; please close the issue.

[matthewevans]

Closing per mike-theDude's triage above — verified fixed/resolved on current main. If this still reproduces on a current build, please reopen with a fresh repro.