fix(settings): preserve distinct catalog identities

[Teakowa]

Summary

Fixes #110 by making settings identity and projection validation fail closed.

• Remove broad key-pattern canonicalization that collapsed distinct hero ability concepts.
• Validate raw hand-written and generated projections before first-wins lookup deduplication.
• Keep hand-written lookup precedence explicit and validate duplicate value shape.
• Add a regression test for distinct hero modifier identities.

Independent QA evidence

This PR was checked against the exact pre-remediation merge head ea45518185c8d34b7186376b9e376e57480b9c85:

• Baseline QA harness: failed as expected because ability1Kb% and ability1SelfKb% both resolved to setting.hero.ability.knockback.
• Treatment QA harness at this branch: passed for distinct ability1Kb%/ability1SelfKb% identities and distinct damageDealt%/damageReceived% identities.
• Disposable mutation: changing generated gamemodes.general.heroLimit from Enum to Bool caused workshop-catalog-gen check to fail with conflicting duplicate settings path.
• The QA harness and mutation worktree were disposable and are not part of this PR.

Local verification

• cargo fmt --all --check
• cargo test --workspace --all-targets --locked
• cargo run -p workshop-rs --bin workshop-catalog-gen -- check
• cargo clippy --workspace --all-targets --locked -- -D warnings
• git diff --check

[Teakowa]

Independent QA audit

Review target: the post-merge findings on #110, not only the implementation-authored tests.

Check	Baseline ea45518	Remediation 62cd832	Verdict
Distinct ability1Kb% vs ability1SelfKb% identity	failed: both setting.hero.ability.knockback	passed	fixed
Distinct damageDealt% vs damageReceived% identity	covered by treatment contract	passed	fixed
Duplicate projection type conflict	not observable through old effective table	disposable mutation rejected by workshop-catalog-gen check	fixed

The baseline/treatment harness and mutation were run outside the committed test surface and then removed. This is an agent-run independent verification pass, not a claim of a separate human approval.

[Teakowa]

Post-remediation QA audit

Blocker

1. Duplicate enum-domain conflicts are still hidden by the validator, and this occurs in the current catalog.

validate_catalog() now compares raw duplicate paths before first-wins deduplication, which is the right direction. However, same_entry_shape() treats every Enum(_) / Enum(_) pair as equivalent and does not compare the enum domain identity.

Current data already contains a concrete mismatch for the same path:

• hand-written gamemodes.general.heroLimit → KeyKind::Enum("heroLimit")
• generated gamemodes.general.heroLimit → KeyKind::Enum("setting_gamemodes_general_heroLimit")

Because both are Enum(_), the raw duplicate check accepts them. entries() then keeps the hand-written entry, so the later SettingDefinition validation never sees the generated conflicting domain. workshop-catalog-gen check is therefore green while the two source projections disagree on the enum domain.

This directly misses #110's acceptance criteria around conflicting domains and stale/duplicate semantic tables. The duplicate comparison needs to compare the full semantic KeyKind (including the exact enum domain), not only the broad variant shape. A regression/mutation should specifically cover Enum("a") vs Enum("b"), not only Enum vs Bool.

Major

1. Duplicate presentation conflicts remain invisible. TableEntry also carries workshop_name, but the raw duplicate reconciliation currently compares only path + broad value shape. Two projections can therefore disagree on the canonical en-US presentation for the same path while check still passes and first-wins lookup silently hides one side. Build the canonical typed Workshop settings catalog #110 explicitly requires stale/duplicate source conflicts and inconsistent presentation/locale facts to fail closed.

The identity-collision remediation itself is valid: the prior contains("Damage") / contains("Kb") style normalization is gone and the reproduced collisions are fixed. CI on 62cd832 is green, but the catalog reconciliation contract is not yet complete.

#110 should remain open until the duplicate projections are compared on the semantic facts that affect the effective catalog rather than only their broad Rust enum variant.