fix(parser): swallow audit — key-anchor the Effect and ActivationRestriction probes

[matthewevans]

Two of the 33 unanchored evidence.any::<T>() probes match a variant name shared
with another tagged enum in a way that CHANGES the detector's answer. Both are now
key-anchored (CR 601.2 / CR 602.5), pinned by differential guard tests observed RED
before green.

REGENERATED THE COLLISION MAP FIRST — the committed one was wrong by omission. It was
computed against 91 enums; the true population is 162 (tag = "type": 93 internally
tagged + 69 adjacently tagged content = "data" — a distinction the old map did not
record, and the two have DIFFERENT deserialize semantics). Same probes, recomputed: 19
colliding (site, variant) pairs across 10 sites, not 13. All 13 previously listed
reproduce; 6 were missed. A collision table is a claim about a population — if the
population is short, its ABSENCES are worthless.

ANCHORED (answer differs across the collision):

any:: CastFromZone also: AbilityCondition, ReplacementCondition
any:: RequiresCondition also: CastingRestriction (field-identical)

Effect::CastFromZone has NINE fields and every one carries #[serde(default)], so it
deserializes from the type tag ALONE — a unit variant in all but syntax. So a cast-zone
CONDITION ("if you cast this spell from your graveyard", CR 601.2a) was read as a one-shot
cast-permission EFFECT and used as proof that the card's "this turn" is intrinsic. A struct
variant is NOT automatically safe: check for a REQUIRED field before believing otherwise.

MEASURED, full 35,396-face pool. The collision surface is POPULATED: 36 CastFromZone
nodes sit under non-effect keys (condition x25, inner x9, conditions x2) and 34
RequiresCondition nodes under casting_restrictions — 70 nodes read as the WRONG enum.
Swallow delta is nevertheless 0 gains / 0 losses (export byte-identical), because both are
MASKED by an upstream text gate, not by any property of the probe: 32 of the 35
CastFromZone cards contain no "this turn" at all, and the activation gate requires every
"this turn" to sit on an "activate only" line, which a casting restriction never does.
That is luck, not safety. Anchoring makes it correct by construction at zero cost.

ANSWER-EQUIVALENT (8 sites / 17 pairs left unanchored, each with a checkable reason):

• the turn-scope disjunction (14 pairs) asks one existential question — "any turn-scoped
  carrier?" — and EVERY variant it matches has ThisTurn in its name. A collision IS name
  identity, so a colliding node still bears a ThisTurn name and is still a turn-scoped
  carrier. The disjunction is closed under its own collision set.
• ManaCost::SelfManaValue vs QuantityRef::SelfManaValue: both ARE a value derived from
  the card's own mana cost (CR 106.1).
• Not: the two condition readings are already OR'd at the call site; FilterProp::Not
  and TargetFilter::Not use different field names, so those collisions are structurally
  impossible.

Anchoring these would be actively WRONG: ReplacementCondition owns YouAttackedThisTurn
and DealtDamageThisTurnBySource and has no leg in that detector, so 16 nodes across ~13
cards are seen today ONLY via the collision — right answer, wrong reason (the
ManaProduction trap from the quantity anchoring). Filed as a follow-up.

GATES (on this base): fmt clean; clippy -D warnings exit 0; 16,281 lib + 2,735 integration
pass, 0 failed; full-pool export BYTE-IDENTICAL vs origin/main — 0 gains / 0 losses over
35,396 faces.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[github-actions]

Parse changes introduced by this PR

✓ No card-parse changes detected.