M19 catalog accuracy and ability-IR scaling audit

[ninthworld]

Tracking issue. Two audits of the M19 catalog and the ability IR, run against the real
set data rather than against the implementation's own tests. Every finding is now a child
issue; this one carries the method, the headline numbers, the sequencing, and the list of
things that were checked and found correct, which lives nowhere else.

Method

Every one of the 299 files in crates/sage-engine/data/catalog/ was matched by name to its
MTGJSON M19.json record (315 printings) and diffed mechanically on mana cost, types,
supertypes, subtypes, power/toughness, loyalty, and colours. Then each of the 267 definitions
carrying an ability block was read side by side with its Oracle text. Every finding that
depends on engine semantics was checked against the code before being reported.

Baseline: main at 02efe72. All of it passes make check.

Children

Part 1 — card text accuracy: 17 of 299 cards are wrong

Issue	What	Scope
#820	Fifteen cards do not match the printed card — 12 wrong mana cost / P/T / subtypes / colours, 8 wrong costs, amounts, or trigger scopes, 5 both. Data only, one PR.	Medium
#821	Two cards the IR cannot currently say — Resplendent Angel is missing an entire ability, Skyrider Patrol asks for two targets where the card names one. Needs vocabulary, not JSON.	Medium
#822	Nothing gates the catalog against real printed data. build.rs checks shape; the behavioural tests assert against whatever the definition says and therefore cement the bug.	Medium

The tests are the sharp end of #822. tests/m19_granted_abilities.rs:169 is named
issue_740_gift_of_paradise_enchants_a_land_and_gains_two_life and asserts life + 2 for a
card that gains three — the name, the assertion, and the comment all encode it, and
docs/generated/test-coverage.md shows the card as covered.

Part 2 — mechanic implementation

The engine is unusually disciplined about purity, and the scripted escape hatch is genuinely
empty — every bundled card is data. The findings are all one problem: the IR grows by naming
phrasings instead of by composing axes, so the cost of the next card is paid in enum variants
and exhaustive match arms rather than in JSON. Effect is at 82 variants against 86 Effect::
arms in sage-server/src/rules_text/effects.rs alone.

The codebase already states the right principle. ObservedActivation's doc comment:

A small product of two independent filters rather than a closed list of named classes, for
the reason PermanentCount is one: the two questions a card asks about an activation … vary
independently, and a variant per pairing would grow the vocabulary once per card.

Every child below is a place where that principle was not applied.

Issue	What	Evidence	Scope
#823	ActivatorScope cannot say "you". Sarkhan's Whelp fires on an opponent's Sarkhan. The gap is not in exclusions.json, whose neighbouring entry asserts the activator filter is complete.	1 card, 0 correct answers available	Small
#824	One permanent filter for the non-targeting selectors. MassAffects, StaticAffects, DestroyAffects, PermanentCount all answer "which permanents?" and none composes. DestroyAffects is two variants for one card and disappears.	9 mass scopes used by exactly 1 card each	Large
#825	Collapse TargetSpec's permanent arms onto that filter. ~29 variants, nearly all a product of axes already modelled.	20 of 36 catalog selectors used by exactly 1 card	Large
#826	One Amount, not a ByCount and a ByAmount twin per verb. 11 variants for 6 verbs, with holes where no M19 card happened to need the pairing.	CreateToken already solved this and its doc says why	Medium
#827	One subject and one modification payload for pump / grant / restrict. The matrix is 9 variants with 4 empty cells — and the empty self-row keyword cell is why #821's Resplendent Angel is missing.	chromium_the_mutable = 3 effects for 1 sentence	Large
#828	Decompose the one-card bundled verbs. sacrifice_chosen_per_player bundles four rules steps and owns the only special case in Effect::target_groups; may_cast_exiled_this_way silently also bottoms a library.	60 of 121 catalog kind values used by exactly 1 card	Large

#825's cost table, for scale — one new TargetSpec variant today touches, all wildcard-free:

Site	Arms
sage-engine/src/resolve.rs — legality	31
sage-engine/src/actions/targeting.rs — candidates	27
sage-server/src/view/prompt.rs — the posed English	29
sage-server/src/rules_text/words.rs	86

Sequencing

#820 ──▶ #822                     data first, then the gate that holds it
  │
  ├──▶ #824 ──▶ #825              selectors: engine half, then the half that
  │                               generates the English a player reads
  ├──▶ #826 ──▶ #827              verbs: amount axis, then subject axis
  │              └── supersedes #821 if scheduled first
  └──▶ #828                       after the above; before another set is authored

#823                              independent, small, any time

• fix(engine): fifteen M19 cards do not match the card that was printed #820 and test(engine): gate the catalog against real printed card data #822 first. Everything in Part 2 re-authors the catalog, and re-authoring wrong
  data preserves the wrongness.
• engine: two M19 cards the ability IR cannot currently say #821 either way. The narrow fix is two affordances; engine: one subject and one modification payload for pump, grant, and restrict #827 supersedes it. Two cards being
  wrong is a worse state to sit in than one field being written twice — AGENTS.md makes
  playing the merge criterion.
• engine: one permanent filter for the non-targeting selectors #824 before engine: collapse TargetSpec's permanent arms onto the shared filter #825, which consumes the filter it introduces.
• engine: one Amount, not a ByCount and a ByAmount twin per verb #826 before engine: one subject and one modification payload for pump, grant, and restrict #827; the payload collapse is much cleaner over verbs that already carry one
  Amount.
• engine: decompose the one-card bundled effect verbs #828 last, and before another set is authored — that is the whole point of it.

#824–#828 are behaviour-preserving refactors. The catalog JSON changes shape, no card's rules
change, and the existing behavioural suites are the safety net for all five.

Related

• Remove exclusions.json and prohibit card/mechanic denylists #791 — remove exclusions.json and prohibit denylists. engine: an observed activation cannot say "you" #823, engine: one permanent filter for the non-targeting selectors #824, engine: collapse TargetSpec's permanent arms onto the shared filter #825, and engine: decompose the one-card bundled effect verbs #828 each drop
  or narrow entries in that file; coordinate rather than duplicate.
• engine: authorable nonsense the validator accepts, and five gaps the M19 tranche left behind #776 — authorable nonsense the validator accepts. Item 7 there is the adjacent half of engine: an observed activation cannot say "you" #823
  (nothing filters an observed activation by whether it was a loyalty ability).

Checked and found correct

Recording these so the next audit does not re-open them. Each looked wrong on a reading of the
card and is right in the code:

Card	Apparent problem	Why it is fine
chaos_wand	no effect bottoms the non-cast cards	MayCastExiledThisWay does it — which is #828's complaint about the verb, not a card bug
fell_specter	player_ref: target_player on a trigger the card does not target with	named_subjects in triggers.rs:587 fixes the discarder from the event
amulet_of_safekeeping	target: spell_on_stack for "counter that spell"	same mechanism, SelfBecomesTarget arm
doublecast	target: spell_on_stack for "copy that spell"	delayed.rs:136 pre-fills the target (CR 603.7c)
skilled_animator	no "for as long as this creature remains" duration	that is Animate::until_end_of_turn's false default
runic_armasaur	no "that isn't a mana ability" filter	structural — a mana ability never reaches the stack (CR 605.3a)
enigma_drake	base power 0 where the card prints *	the catalog convention: base 0 plus defined_power, reading the graveyard

Also verified clean: all 14 keyword abilities across all 299 cards match the printed keyword
line, with exactly one exception (mistcaller's phantom flash, in #820); and every card the
catalog treats as vanilla really is vanilla.