Clear stdlib ABC caches in purge_modules (fixes flaky 3.10 suite)

[JarryShaw]

Fixes a latent test-harness bug that surfaces as intermittent, order-dependent failures on Python 3.10. It is what blocks CI on #378, but it is not an ESP bug — it predates that branch and affects the whole suite.

Symptom

On 3.10, the full suite intermittently fails one of:

• test_context_registry_normalisation → RegistryError: not a protocol context: 'ESP'
• test_*_option_constructors_cover_* (ipv4/mh/ipv6) → AttributeError: 'dict' object has no attribute 'to_dict'

Every one of these passes when its file runs alone. Which test fails depends on collection order.

Root cause

tests/_support.purge_modules(['pcapkit']) — called in almost every setUp — drops pcapkit from sys.modules so the next test re-imports it fresh. But the stdlib collections.abc ABCs are never purged. Each re-import rebuilds pcapkit's Mapping subclasses (Info, Schema, ContextRegistry, ProtocolContext, EnumSchema) as brand-new class objects, and that repeated subclass churn corrupts the C-level _abc_impl instance-check caches on the shared ABCs.

The caches then return stale answers for immortal built-ins:

• isinstance({}, collections.abc.Mapping) → False, so ContextRegistry.make({'ESP': ctx}) skips its Mapping branch, iterates the dict's keys, and calls register('ESP').
• isinstance({}, Schema) → True, so Schema.to_dict() (schema.py:437) is called on a plain dict.

It self-heals by teardown as the cache token advances, which is exactly why it is invisible per-file and only bites the full suite.

Confirmed pre-existing: the commit before the ESP const/vendor work fails identically, and pcapkit/corekit/infoclass.py:270 already carries a defensive isinstance(dict_, (dict, collections.abc.Mapping)) — a prior workaround for the same fragility in one spot.

Fix

purge_modules now resets the ABC caches after the purge, via abc._reset_caches over the collections.abc members. It is a CPython internal (present on the C _abc and pure-python _py_abc backends), guarded with getattr so a future runtime without it degrades to the old occasionally-flaky behaviour rather than erroring. gc.collect() was tried and does not work — the poison is in the ABC caches, not in dead subclasses.

Verification (Python 3.10)

• Fix branch CI selection: 347 passed, 14 skipped, 156 subtests, 0 failed.
• Applied onto Implement ESP, with optional payload decryption #378's branch, its previously-failing CI selection: 328 passed, 27 skipped, 118 subtests, 0 failed (was 1 failed).
• tests/protocols/internet alone on Implement ESP, with optional payload decryption #378's branch: 88 passed, 14 skipped, 0 failed (the to_dict cascade gone).

#378 will be rebased onto this once it merges.

[Copilot]

🟡 Changes recommended

The new _reset_abc_caches() implementation calls abc._reset_caches with an argument, which is incompatible with CPython’s no-arg internal API and will raise TypeError during test setup.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR addresses intermittent, order-dependent test failures on Python 3.10 by strengthening the test harness’s module-purge behavior to avoid stale collections.abc instance-check cache results across repeated pcapkit re-imports.

Changes:

• Add a helper to reset stdlib ABC caches after pcapkit is purged from sys.modules.
• Invoke the ABC-cache reset from purge_modules to reduce cross-test contamination.

File summaries

File	Description
tests/_support.py	Adds _reset_abc_caches() and calls it from purge_modules() to mitigate flaky failures caused by stale ABC instance-check caches across repeated re-imports.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟢 Approval recommended

The change is test-harness-only, uses guarded internal APIs with a safe fallback, and directly targets the documented Python 3.10 flakiness mechanism.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0 new
• Review effort level: Lite

[JarryShaw]

Reverted the tests: handle abc cache reset API variants commit (force-pushed back to d0def088) — it broke the whole suite rather than hardening it, and the concern it addressed does not hold.

abc._reset_caches has a stable signature across CPython versions: it takes exactly one argument, the class. There is no no-arg form. So the amended code's reset() raises TypeError on every supported Python (3.10 and 3.14 both confirmed: _abc._reset_caches() takes exactly one argument (0 given)), which is swallowed, and control falls to the _abc_caches_clear() fallback. That fallback is what breaks CI: on 3.13/3.14 Mapping._abc_caches_clear is an unbound method requiring cls, so calling it with no argument raises TypeError: ABCMeta._abc_caches_clear() missing 1 required positional argument: 'cls' — and that raise is not inside the try, so it propagates straight out of purge_modules. Since almost every test's setUp calls purge_modules, every one of them errors — hence "all failing".

The original per-class form (abc._reset_caches(obj) for each collections.abc ABC) is the correct call and is already guarded with getattr(abc, '_reset_caches', None), which degrades gracefully if a future runtime ever drops the helper. Verified green on 3.10 (347 passed / 14 skipped / 0 failed) and applied onto #378's branch (328 passed / 27 skipped / 0 failed).