Unbounded cache in SpecValidator.iter_errors retains validator instances and schemas

Start in openapi_spec_validator/validation/validators.py around SpecValidator.iter_errors at lines 79-86, and run the minimal reproduction to confirm cache growth and release after cache_clear(). Trace how validate() creates validators and how repeated iteration is preserved. Done means discarded specifications and validator instances can be collected without unbounded historical retention, while repeated iteration still works.

Bug description

Repeated calls to validate() permanently retain each validator instance and its schema, even after validation succeeds, the caller discards the schema, and gc.collect() runs. This causes approximately linear memory growth in long-running applications that repeatedly load/validate specifications.

Reproduced with openapi-spec-validator 0.9.0, CPython 3.11.15, macOS arm64. The same retention mechanism was also reproduced with 0.8.5. The reproduction below uses only this package and the Python standard library; no Prance, web server, application code, or external data is needed.

Minimal reproduction

Install openapi-spec-validator==0.9.0 and run in a fresh process:

import gc
import tracemalloc
from importlib.metadata import version

from openapi_spec_validator import validate
from openapi_spec_validator.validation.validators import SpecValidator


def make_spec():
    return {
        'openapi': '3.0.0',
        'info': {'title': 'Memory reproduction', 'version': '1.0.0'},
        'paths': {},
        'x-payload': 'x' * 65536,
    }


# Warm up lazy imports; clearing the private cache is diagnostic only.
validate(make_spec())
cache = SpecValidator.iter_errors.__wrapped__
cache.cache_clear()
gc.collect()
tracemalloc.start()
print('openapi-spec-validator', version('openapi-spec-validator'))
for count in (50, 100, 150):
    for _ in range(50):
        validate(make_spec())
    gc.collect()
    print(count, 'cached validators:', cache.cache_info().currsize,
          'retained bytes:', tracemalloc.get_traced_memory()[0])
cache.cache_clear()
gc.collect()
print('after cache_clear:', tracemalloc.get_traced_memory()[0])
tracemalloc.stop()

Observed output (exact byte counts may vary):

openapi-spec-validator 0.9.0
50 cached validators: 50 retained bytes: 3498052
100 cached validators: 100 retained bytes: 6960377
150 cached validators: 150 retained bytes: 10417844
after cache_clear: 23183

Each input is a freshly allocated but content-identical valid specification, containing a synthetic 64 KiB extension value. No input or validation result is retained by the caller. The memory values are current Python allocations measured by tracemalloc after GC, not peak RSS.

Expected behavior

Once validation finishes and the caller releases the specification, the validator and its schema should be eligible for garbage collection. Repeated validation should not create an unbounded process-wide collection of historical validator instances.

Suspected cause

SpecValidator.iter_errors is decorated with lru_cache(maxsize=None). Since the cache key includes self, every instance created by the validate() shortcut remains reachable from the class-level cache, together with its schema. This happens even for successful validations with no errors.

Clearing this private cache releases almost all of the retained memory in the experiment. The __wrapped__/cache_clear() calls above are diagnostic controls, not a proposed application workaround: globally clearing the cache can interfere with unrelated callers.

Could the cached iterable be owned by the validator instance, or otherwise avoid globally retaining validator instances, while preserving repeated iteration behavior?

I searched existing issues/PRs for memory, leak, lru_cache, and iter_errors but did not find an equivalent report.