Index literal union members by value instead of scanning pairwise

[JulianaSaran]

Description

A condition with a big pasted literal list (BusinessActivities intersects ['DRI-2090', … ×940])
compiles quadratically, twice: UnionType::join dedups members pairwise (recomputing both shapes
per pair), and UnionShape::of scans pairwise again on every shape() call — one intersects
resolve asks ~4 times. ~880k comparisons per condition; a real product with 34 endorsement
conditions takes 14.2s to compile, so production hits the 30s limit rendering the
publication-checks badge (whose cache can then never warm) and running quote previews. Prod
traces die in each of the two loops (attached as a comment).

Literals are now found through an index keyed by LiteralShape::valueKey() — a hash key
consistent with equals: equal literals always share a key; sharing one only nominates, the
sameness judgment still decides (big ints sharing a float image stay distinct, NAN equals
nothing). Non-literal members keep the pairwise scan — only literals arrive in bulk. Same move
as #73; here the key isn't the whole equality, so the bucket verifies.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• Performance improvement

Related Issues

Relates to IP-475 (internal tracker; production timeouts in Product Catalogue).

Changes Made

• LiteralShape::valueKey() — canonical hash key consistent with value equality.
• DistinctShapes (@internal) — insert-only set under an injected sameness judgment;
  literals via the key index, everything else pairwise.
• UnionShape::of and UnionType::join dedup through it; join computes each type's
  shape once instead of once per comparison.
• Tests: a pairwise oracle (of() ≡ naive scan over a hostile pool) and a lockstep law
  (ValueEquality::equals(a,b) ⇒ same key, with drift sentinels: case, whitespace,
  unicode composition, ±0.0, the 2^53 boundary, NAN/INF), plus corner pins.

Testing

• All existing tests pass (composer test) — 1,316 tests, zend.assertions=1
• Added new tests for the changes
• Static analysis passes (composer test:types)
• Code coverage remains at 100% — no coverage driver on my machine; relying on CI
• Mutation testing passes (composer test:infection) — same, relying on CI

Code Quality Checklist

• My code follows the project's coding standards
• I have run vendor/bin/pint to format my code
• I have added/updated PHPDoc comments where necessary
• I have updated the documentation (if applicable)
• My changes generate no new warnings or errors
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Documentation

• I have updated the README.md (not applicable — internal performance change)
• I have updated the CHANGELOG.md
• I have added/updated code examples (not applicable)

Breaking Changes

N/A — behaviour is pinned identical by the oracle test; the index only changes cost.

Additional Notes

• Bulk non-literal unions (e.g. hundreds of record literals) stay pairwise — same as
  today, no regression.
• UnionShape::equals and union×union overlap remain O(N·M) — follow-up material.
• Deeper option: a value-keyed set representation for enum-sized unions (or full
  hash-consing) — this PR is the minimal step and doesn't block it; the oracle and
  lockstep tests carry over unchanged.

🤖 Generated with Claude Code