[Volv-G]

Summary

validate_dehydrated_data rebuilt the Draft 2020-12 validator and re-ran check_schema() against the constant packaged dehydrated schema on every call. This PR builds the validator once.

Design (minimal)

• New private no-arg factory _dehydrated_validator() decorated with @lru_cache(maxsize=1): loads the (already lru_cached) packaged schema, runs validator_for(schema).check_schema(schema) once, and returns the constructed validator.
• validate_dehydrated_data calls the factory instead of rebuilding per call. No locks, copies, or global state beyond the lru_cache.
• Only the validator is cached — never validation results. Every payload is still fully validated.

Semantics

• Validation semantics and errors are unchanged: same iter_errors → sort by absolute_path → best_match path, same SchemaValidationError message and location.
• An invalid schema would still raise SchemaError on every call (lru_cache does not cache exceptions).

Tests

• test_check_schema_runs_once_across_many_validations: 125 interleaved valid/invalid validations → check_schema called exactly once (counted via monkeypatched Draft202012Validator.check_schema, cache cleared before/after each test).
• test_cached_validator_keeps_error_messages: on a warm cache, each invalid payload's message + location matches the uncached reference path exactly, with valid payloads interleaved.
• Results: 41 passed in tests/test_schema_validation.py; full suite 2434 passed.

Benchmark (old per-call path vs cached, min of 5×300 calls)

payload	before	after	speedup
valid	36.3 ms/call	0.67 ms/call	54×
invalid	22.7 ms/call	0.73 ms/call	31×

Residual note

The cached validator shares the packaged schema dict returned by load_dehydrated_schema(). Nothing mutates it today; if code ever mutated that dict in place or introduced a custom/swapped schema, it would need to call _dehydrated_validator.cache_clear() to re-run check_schema/rebuild (tests do this explicitly).