fix: Report each composition branch error once, with a structured error - #53

Closes #52.

Composition branch errors were lost, duplicated, or inconsistent between
keywords, and getErrors() could come back empty. Three defects, all in the
branch-error plumbing shared by allOf / anyOf / oneOf:

1. Empty error list. AbstractCompositionalValidator::validateBranch()
   caught the InvalidDataTypeException thrown by SchemaValueNormalizer and
   wrapped it in a bare ValidationException — no errors:, no
   abstractErrors. Nothing structured was ever synthesised for the rejected
   value, so getErrors() was [] and getFormattedErrors() was "".
   OneOfValidatorWithContext::validateWithoutDiscriminator() reached the same
   end state via catch (InvalidDataTypeException) { continue; }.
2. allOf duplication. validateBranch() returned each failure twice — as
   errors: [$e] and again as abstractErrors extracted from
   $e->getErrors() — and AllOfValidator merged both buckets.
   AnyOf/OneOfValidator read only abstractErrors, hence the disagreement
   between keywords.
3. Miscounted failures. The allOf message formatted
   count($result->errors), which a branch throwing a bare
   AbstractValidationError never reached — "but 0 failed" beside a
   non-empty error list.

Changes

• BranchOutcome and ValidationResult now carry a single canonical error
  list (list<ValidationErrorInterface>) instead of two overlapping ones, so
  no caller can double-count. ValidationResult::$abstractErrors is merged
  into $errors, and a new $failedCount records how many branches did not
  match.
• Where InvalidDataTypeException was swallowed, both validator families now
  synthesise a TypeMismatchError carrying the branch's dataPath and
  schemaPath (/allOf/0, /oneOf/1, …) with actual: 'null' for the common
  case — following the DependentSchemasValidator precedent of never throwing
  with an empty list. A ValidationException from a branch that carries no
  structured errors falls back to NestedValidationError for the same reason.
• The allOf message derives its count from failedCount, not from an error
  bucket.
• MAX_COMPOSITION_ERRORS / TooManyErrorsError now caps the merged list.
  Behaviour is equivalent for the previously-counted case (20 branch errors →
  20 + TooManyErrorsError); previously allOf re-appended the wrapper
  errors after the cap, so a 30-failing-branch allOf returned 41 errors
  despite the cap. It now returns 21. Covered by
  CompositionBranchErrorsTest::composition_errors_are_capped.

Error identity is preserved: nested branch failures keep their original
keyword / dataPath / schemaPath rather than being flattened into a generic
wrapper. DiscriminatorDataError and OneOfError reporting are untouched.

Before / after

validateSchema(), running the repro from #52 verbatim
(B = FormIdentifier, errors= is count($e->getErrors())):

Message wording for anyOf / oneOf is unchanged.

PSR-7 validateResponse(), root-level content.application/json.schema of
{allOf: [$ref B]}:

And {allOf: [B, C]} with both branches failing now reports one entry per
distinct violation with "but 2 failed" (was 4 entries).

Tests

Written before the fix, each verified failing against main first:

• tests/Unit/Validator/SchemaValidator/CompositionBranchErrorsTest.php —
  AllOf/AnyOf/OneOfValidator (the non-context family): per-branch
  schemaPath indices, single-branch and multi-branch failure counts,
  multiple errors from one branch, keyword agreement, and the error cap.
• tests/Integration/Validator/Schema/OneOfValidatorWithContextTest.php —
  OneOfValidatorWithContext: structured errors for rejected null
  (including untyped, scalar, and type array branches), no duplication,
  accumulation across branches, and preservation of every error from a single
  branch.
• tests/Functional/Schema/CompositionBranchErrorsTest.php — the Composition validators lose, duplicate, and disagree about branch errors; getErrors() can be empty #52 repro
  through validateSchema(), plus getFormattedErrors() is never empty and
  never repeats.
• tests/Functional/Response/CompositionBranchErrorsTest.php — the PSR-7
  validateResponse() path with a null body.

ValidationResultTest is updated for the merged bucket.

make tests (7161 tests, green — the 2 reported deprecations are pre-existing
and unrelated), make psalm (no errors), make cs-fix, make rector all
clean. make infection scoped to the seven touched files: covered MSI 85% →
91% (gate is 78%). The remaining escaped mutants are on lines this PR does
not modify, plus one harmless array_values() unwrap.

Notes

• Out of scope, per the issue's "Related observation": which of several
  violations surfaces first still depends on the wrapper — branches go through
  the non-context SchemaValidator, which orders keyword groups differently
  from SchemaValidatorWithContext. This PR does not change that (see the
  oneOf + {type: widgets} row: required@/ before and after), and neither
  path is exhaustive. The one functional assertion that would have straddled
  the two families is scoped to allOf/anyOf with a comment pointing at Composition validators lose, duplicate, and disagree about branch errors; getErrors() can be empty #52.
• ValidationResult is a shared internal DTO of the compositional validators
  (its only callers are the four classes touched here); its property set
  changes as described above. ValidationException::getErrors() and the
  AbstractValidationError shape are unchanged — this PR only adds errors
  where there were none.
• Independent of fix: Honor nullable when it sits beside a composition keyword #51 (issue Combining "nullable" and "allOf" does not work correctly #50), which touches the same three validate()
  methods but only adds an early return at the top of each. This branch is
  cut from main; it will need no more than a trivial rebase whichever lands
  first.