refactor(parser): Plan 05b T8-A4 — U0-53, the one R2 row under the cap

[matthewevans]

Plan 05b tranche T8-A4 — the R2 rows, and the last sub-tranche of T8. Follows #6719 (A3).

One row converted, four excluded with evidence. The exclusion table is this tranche's headline
deliverable, not the conversion count.

Why the scope is 1 of 5

§T8 lists five R2 rows (U0-53, U0-62, U0-02, U0-52, U0-38) but its own R2 paragraph caps R2 at
single-clause bodies with no sub_ability, "anything larger is a recognizer rewrite, not an R2
conversion". Neither statement wins outright, and the four exclusions have three different
reasons — for two of them the cap is not even the operative axis.

row	site	cap verdict	action
U0-53	oracle_class.rs:118-172	under cap — 1 clause, no sub_ability	CONVERTED
U0-38	oracle.rs:5511-5531	exceeds cap	excluded
U0-52	oracle.rs:6392	not an R2 site at all	excluded
U0-02	oracle.rs:2164	passes structural cap, fails §5.2	excluded
U0-62	oracle_class.rs:341-344	passes structural cap, fails §5.2	excluded

U0-38 exceeds the cap textually. oracle.rs:5519 is
.sub_ability(AbilityDefinition::new(AbilityKind::Spell, Effect::PreventDamage{..})) — a two-node
Choose → PreventDamage body. Corroborating from a second direction: the recensus's own tranche
table at line 601 assigns U0-38 (and U0-02) to T10e, not T8. Two of the document's three
statements agree; the §T8 row list is the outlier.

U0-52 is not a hand-built-body site. oracle.rs:6382 binds nom_def once; U0-51 emits it when
the dispatcher produced a real effect, and U0-52 emits the same binding otherwise. The five
Effect::unimplemented shapes the census attributes to this site are produced inside
dispatch_line_nom. Converting it needs an AbilityIr for the fallback branch, and there are exactly
three ways to get one: an IR-returning dispatcher (precisely T10a's stated blocker), re-running the
dispatcher (the census forbids it), or un-lowering the definition (impossible —
AbilityIr::from_definition does not exist, by A0's H4 design).

U0-02 and U0-62: verdict confirmed, reasoning corrected. Both bodies are make_unimplemented(...)
— single clause, no sub_ability — so they pass the structural cap, and the implied byte-identity
blocker is not real: Effect::unimplemented("unknown", line) expands to exactly that literal, so
Gate F is satisfiable and byte-identity is mechanically reachable. Gate F is not the blocker.

The verdict holds on firmer ground. §5.2 makes the written CR-faithfulness argument a mandatory half
of the gate, and requires it "for the CLASS of text the recognizer handles". Neither site has a
recognizer. Both are terminal residuals whose membership is defined negatively — whatever every
preceding arm declined. A negatively-defined set has no text class, so the required argument cannot be
written for either site under any recipe. That is a structural property, not a judgement call, and
it survives even if the shared residual node is later built. Wrapping the sentinel ad hoc at two sites
would also be the exact per-site special case that node exists to replace.

The shared residual node was not built — architecture work, out of scope.

The converted row: U0-53, the Class level bar

76 of 79 Class abilities. kind = Activated, effect = SetClassLevel{level}, plus cost,
description, and activation_restrictions = [AsSorcery, ClassLevelIs{level-1}] in that order.

Labelled correct-by-corpus, and R1's "no property of any card's text participates" is deliberately
not borrowed. What makes it byte-identical:

• Effect::SetClassLevel occurs exactly once in the whole of parser/oracle_effect/ —
  sequence.rs:6175, inside clause_is_dig_lookback_transparent, a predicate that runs only when an
  antecedent search walks back over intervening clauses. A one-clause chain has none, so it is never
  invoked. Every other pass in the assembly tail and in finalize_effect_chain's six passes keys on a
  different Effect variant, or on sub_ability/else_ability/mode_abilities presence — none of
  which this body has.
• PlayerScopeRewrite::Preserve removes apply_player_scope_rewrites by construction.
• The owner-library anchor is doubly inert: its guard needs two phrases the level line cannot contain,
  and the walk beneath it looks for Effect::Shuffle/Effect::RevealTop, neither expressible here.

What a future card could violate: a Class card printing "Level 0" — section.level - 1 on a u8
underflows. Pre-existing, preserved, not introduced.

What a future commit could violate — the honest limit of R2. Unlike R1's algebraic identity, this
argument is a survey of today's pass list. A pass added later to finalize_effect_chain or
assemble_effect_chain's tail that keys on AbilityKind::Activated, on chain length, or on
SetClassLevel would newly apply here, with nothing in the type system to catch it. This is why R2
is not correct-by-construction, stated rather than glossed.

§5.3

Reachability, attributed by probe message and never by test name: 21 reachers — 9 in --lib, 12 in
the integration binary. Attribution mattered a third consecutive time: bandits_talent_level3_… and
all ten Scavengers-Talent / Party-Dude integration tests name a card and never the recognizer.

No remediation snapshot was needed — the first tranche in A1–A4 where that is true. The five Class
two-layer snapshots landed at T1 already witness this site twice each (Level 2 and Level 3), across
ten distinct cost shapes. Their *_ir.snap payloads were checked to confirm they serialize the full
lowered def including the ordered activation_restrictions array — that is why the perturbations
discriminate, verified rather than assumed.

perturbation	result
swap restriction order (ClassLevelIs before AsSorcery)	RED — 6 tests
drop shell description	RED — 6 tests
drop shell cost	RED — 6 tests
PlayerScopeRewrite::Apply instead of Preserve	NULL — 137/137 pass either way

The Apply/Preserve null deserves emphasis rather than burial. Preserve is one of §T8's two
mandated R2 mitigations, and measured, it is extensionally inert on today's corpus — so it supplies
no evidence that this conversion is safe. Preserve remains the correct value (it is what the
hand-built path did, by construction), but the safety argument rests entirely on the shape analysis
above. This row should not be read as "mitigation applied, therefore safe".

ChainLoweringMode is inapplicable here rather than null: the recognizer parses no text, it
hand-builds an Effect from a parsed level number, so there is no entry point whose mode could be
crossed.

Gates

cargo fmt --all                                      clean
cargo clippy -p engine --lib -- -D warnings          zero warnings
cargo nextest run -p engine --lib                    17851 passed, 6 skipped
cargo nextest run -p engine --test integration       4123 passed, 2 skipped
./scripts/check-prelowered-ratchet.sh                Gate P PASS
./scripts/check-skill-doc.sh                         ✓ oracle-parser skill references valid
./scripts/check-parser-combinators.sh                Gate G PASS / Gate A PASS

Zero snapshot churn: the entire diff is one line — M crates/engine/src/parser/oracle_class.rs.
No files added, no snapshots changed, .snap.new count 0.

Gate P did not lower for oracle_class.rs, deliberately. It stays at 5/5, ledger unedited. That
file has no DocEmitter, so there is no ability_ir_at seam to hide the token behind; the site now
states phase A's identity inline as PreLoweredSpell(lower_ability_ir(&ir)) — the same value
ability_ir_at computes one file over. The token retires at T9 with the payload swap.

Harvest (§5.4)

1. oracle_class.rs cited CR 716.4 for the Class level gate — the rule that explicitly disclaims the
   connection. CR 716.4 (:6030) is about leveler cards and states they "are not the same as
   class level abilities. Level counters do not interact with Class cards, and class levels do not
   interact with leveler cards." The annotated claim is CR 716.2a's (:6020). CR 602.5d was
   correct and is kept. Fixed in this commit. The U0-53 census row carries the same wrong
   number — the recensus needs correcting, which is a maintainer call; no planning doc was edited.
2. Restriction order is the reverse of CR 716.2a's stated order ("only if this Class is level N-1
   and only as a sorcery"). Directly analogous to A2's Boast/CR 702.142a finding. Semantically
   harmless — the restrictions are a conjunction — and preserved, not fixed.
3. section.level - 1 is an unguarded u8 subtraction; a "Level 0" line would underflow.
   Unreachable today only as a corpus property. Pre-existing.
4. The SetClassLevel single-occurrence grep is reusable as the mechanical core of any future
   single-leaf-Effect R2 argument.
5. U0-62's residual is reached by two routes, one of which DISCARDS a parsed chain.
   oracle_class.rs:332-338: is_effect_sentence_candidate → parse_effect_chain → if
   has_unimplemented, the parsed def is thrown away and the line falls through to
   make_unimplemented(line). Whoever designs the shared residual node must decide whether to preserve
   that discarded partial chain — and doing so is a behaviour change, not a conversion.
6. U0-53 was the last hand-built AbilityDefinition body in oracle_class.rs — its
   AbilityDefinition import is now unused and was removed.

Deviation flagged rather than buried

A pre-existing CR citation was corrected inside a byte-identity commit. Justified by CLAUDE.md
("if you are modifying existing game logic, verify existing CR annotations are present and still
accurate"); it changes no code, and leaving a known-wrong CR number in a block being rewritten was the
worse option. Flagged because byte-identity tranches are otherwise strictly no-extras.

Summary by CodeRabbit

• Bug Fixes
  • Improved Oracle class level-up ability parsing and processing.
  • Ensured level-up effects are consistently converted into the game’s ability system, including activation costs and class-level restrictions.
  • Preserved player scope behavior when applying level-up effects.

[coderabbitai]

📝 Walkthrough

Walkthrough

The class parser now constructs level-up abilities as effect-chain IR, preserves player scope, and lowers the resulting AbilityIr through lower_ability_ir instead of directly emitting an AbilityDefinition.

Changes

Oracle class level-up parsing

Layer / File(s)	Summary
Construct and lower level-up ability IR crates/engine/src/parser/oracle_class.rs	The {cost}: Level N path now builds an EffectChainIr containing Effect::SetClassLevel, wraps it in an AbilityIr with the parsed cost and activation restrictions, preserves player scope, and emits the lowered result through lower_ability_ir.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• phase-rs/phase#6709: Updates adjacent class-level parsing to emit trigger IR through TriggerNodeIr::from_definition.
• phase-rs/phase#6714: Introduces related shell-widening changes used by lower_ability_ir and AbilityShellIr.
• phase-rs/phase#6715: Refactors Oracle parser ability generation to use AbilityIr and effect-chain lowering.

Suggested labels: quality

Suggested reviewers: nghetienhiep

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is related to the parser refactor and correctly points to the U0-53 class-level row being changed.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch ship/plan05b-t8a4

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/engine/src/parser/oracle_class.rs`:
- Around line 161-163: Update parse_class_level_line when constructing
ActivationRestriction::ClassLevelIs to reject or handle section.level == 0
before evaluating section.level - 1. Preserve the existing level subtraction
only for valid positive levels, preventing underflow on malformed class text.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: 8d5acc5b-be16-4c95-99d7-543a5035e001

📥 Commits

Reviewing files that changed from the base of the PR and between f546654 and 1eb4702.

📒 Files selected for processing (1)

• crates/engine/src/parser/oracle_class.rs

[coderabbitai]

🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Confirm parse_class_level_line's return range for `level`
ast-grep run --pattern 'fn parse_class_level_line($$$) { $$$ }' --lang rust crates/engine/src/parser/oracle_class.rs

Repository: phase-rs/phase

Length of output: 152

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Map the relevant file and find the parsing helper/uses.
ast-grep outline crates/engine/src/parser/oracle_class.rs --view expanded || true
rg -n "parse_class_level_line|section\.level|ClassLevelIs|level - 1" crates/engine/src/parser/oracle_class.rs crates/engine/src/parser -S

Repository: phase-rs/phase

Length of output: 8507

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

# Read the relevant section with line numbers.
sed -n '1,260p' crates/engine/src/parser/oracle_class.rs | cat -n

Repository: phase-rs/phase

Length of output: 13999

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

python3 - <<'PY'
from pathlib import Path
p = Path('crates/engine/src/parser/oracle_class.rs')
text = p.read_text()
for needle in ['parse_class_level_line', 'section.level', 'ClassLevelIs', 'level - 1']:
    print(f"\n## {needle}")
    for i, line in enumerate(text.splitlines(), 1):
        if needle in line:
            print(f"{i}: {line}")
PY

Repository: phase-rs/phase

Length of output: 1669

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

echo "### oracle_class.rs matches"
rg -n "parse_class_level_line|section\.level|ClassLevelIs|Level [0-9]+" crates/engine/src/parser/oracle_class.rs -S || true

echo
echo "### oracle_class.rs excerpt"
nl -ba crates/engine/src/parser/oracle_class.rs | sed -n '1,260p'

Repository: phase-rs/phase

Length of output: 1873

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail
nl -ba crates/engine/src/parser/oracle_class.rs | sed -n '120,210p'

Repository: phase-rs/phase

Length of output: 192

---

🌐 Web query:

CR 716.2a level up first level level 2 class cards level 1 class abilities

💡 Result:

In Magic: The Gathering, rule 716.2a defines how the activated abilities on Class enchantments function [1][2]. Specifically, the rule states that an ability formatted as "[Cost]: Level N — [Abilities]" means "Activate only if this Class is level N-1 and only as a sorcery" [1][2]. Class cards operate according to the following mechanics: 1. Starting State: When a Class enchantment enters the battlefield, it is at level 1 and has access to its first (level 1) class ability [3][4]. 2. Leveling Up: To advance a Class to the next level, you must activate the corresponding level-up ability by paying its cost [4][5]. This is an activated ability that uses the stack and can be interacted with (e.g., countered) [5]. 3. Timing and Requirements: Per rule 716.2a, you can only activate a level-up ability at sorcery speed, and only if the Class is currently at the level immediately preceding the level you are attempting to reach (e.g., you must be at level 1 to activate the ability to reach level 2) [3][6][4]. 4. Cumulative Abilities: Gaining a level does not remove abilities the Class had at previous levels [7]. When a Class reaches a new level, it retains all its previous class abilities and gains the new one associated with the new level [3][4][7]. Each Class card typically contains five abilities: three "class abilities" (which represent the effects gained at levels 1, 2, and 3) and two "level abilities" (the activated abilities used to advance from level 1 to 2 and from level 2 to 3) [3][6].

Citations:

• 1: https://www.reddit.com/r/mtgrules/comments/1gohi4m/question_about_bello_marvin_and_alchemists_talent/
• 2: https://www.reddit.com/r/mtgrules/comments/16d6sko/agathas_soul_cauldron_and_level_up_cards/
• 3: https://gatherer.wizards.com/Pages/Card/Details.aspx?name=Monk%20Class
• 4: https://cardgamebase.com/mtg-class/
• 5: https://draftsim.com/mtg-class-cards/
• 6: https://scryfall.com/card/pafr/202a/ranger-class
• 7: https://scryfall.com/card/pafr/228p/monk-class

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail
nl -ba crates/engine/src/parser/oracle_class.rs | sed -n '140,190p'

Repository: phase-rs/phase

Length of output: 192

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

rg -n "fn parse_number|parse_number\\(" crates/engine/src/parser -S

Repository: phase-rs/phase

Length of output: 29357

---

🏁 Script executed:

#!/bin/bash
set -euo pipefail

echo "### parse_number implementation"
sed -n '1,140p' crates/engine/src/parser/oracle_nom/primitives.rs

echo
echo "### parse_number tests around zero / digits"
sed -n '1400,1675p' crates/engine/src/parser/oracle_nom/primitives.rs

echo
echo "### oracle_class tests around level lines"
sed -n '460,580p' crates/engine/src/parser/oracle_class.rs

Repository: phase-rs/phase

Length of output: 21537

---

Reject Level 0 before subtracting 1. parse_class_level_line accepts Level 0, so section.level - 1 can underflow on malformed class text; gate it explicitly when building ClassLevelIs.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/engine/src/parser/oracle_class.rs` around lines 161 - 163, Update
parse_class_level_line when constructing ActivationRestriction::ClassLevelIs to
reject or handle section.level == 0 before evaluating section.level - 1.
Preserve the existing level subtraction only for valid positive levels,
preventing underflow on malformed class text.

[github-actions]

Parse changes introduced by this PR

✓ No card-parse changes detected.