test: keep the per-class pcapkit import in 12 classes that re-imported per test (#1538)

[JarryShaw]

Please follow the guide below

• You will be asked some questions, please read them carefully and answer honestly

• Put an x into all the boxes [ ] relevant to your pull request (like that [x])

• Use Preview tab to see how your pull request will actually look like

• Searched for similar pull requests

• Followed the coding style (make pylint, make mypy, make isort)

• make test passes, and a test case covers the change (affected legs only, listed below; the full suite was not run locally)

• Added a changelog entry under docs/source/changelog/ and regenerated CHANGELOG.md, if the change is user-visible: N/A — centralised in docs(changelog): shared 1.5.0 changelog — long-lived, merges last (#610, #616, #617, #618, #620) #657

What is the purpose of your pull request?

Tick the commit type your subject line carries.

• fix — corrects a defect
• feat — adds a feature
• perf — changes performance, not behaviour
• refactor — changes neither behaviour nor performance
• test — tests only
• docs — documentation only
• ci — workflows or build tooling
• release — bumps the version or rolls up a distribution
• chore — anything else

---

Description of your pull request and other information

Refs #1538 (P2). Twelve classes in seven modules, 66 tests, called reimport_once_per_class() in setUp and then purged pcapkit in tearDown. That emptied the class's import before _keep_lazy_imports could record it, so every test re-imported the package. This PR drops the tearDown purge. The proposal's 13 classes and 72 tests counts are wrong: an AST scan of main finds 12 and 66. test_const_enum_builtin_parity is untouched.

• Schema ABC cache and default dispatch: both now call restore=True, so their documented confinement (the ABC-cache reset, the reloaded Raw) holds at class scope. Mutation tests give the same failing tests before and after: reverting Schema subclasses share one ABCMeta cache on Python <=3.10, so isinstance/issubclass return whichever answer was asked first #439 (type.__new__) fails the same 5, and a pre-Memoise the default ModuleDescriptor on a registry miss, instead of re-resolving it per frame (from #563) #574 klass fails the same 1. A perf: Optimize Protocol extraction by caching dynamically imported modules in ProtocolBase #563-style memo fails 2 here against 1 on main.
• Found a latent order dependence: a dispatch hit writes the resolved class back into the registry. So test_dispatch_keys_did_not_move_with_the_modules failed on the shared import whenever a parsing test ran first: reverse order gave 2 failures, and so did 3 of 3 shuffles. The per-test purge had been hiding this. That test now takes isolate_modules(self), and 17 orderings pass.
• Guard: tests/project/test_per_class_import_guard.py fails on main, listing the 12 classes. It has an allowlist (empty) that is enforced both ways, and it does not match narrow purges such as the test_scapy_unit one.

Measured, serial pytest, mean of 3:

Run	Before	After
The 7 modules in one process	71.8 s	15.4 s
Sum of 7 per-module processes	72.0 s	22.6 s

Both runs gave 68 passed and 63 subtests. Line and branch coverage of pcapkit is identical (27,861 lines, 1,307 branches).

Plain unittest, one run each:

Leg	Before	After
run_unittest_leg.py protocols --exclude protocols/internet	330 s	286 s
run_unittest_leg.py utilities	63 s	48 s

Both legs had 0 failures and 0 errors. These figures are lower bounds, because the baseline was a non-git tree that skipped 57 git-dependent root tests.

CI estimate, not measured: about 2.4-3.7 runner-minutes per PR, using #1538's unverified 1.4-2.2x xdist factor and its x1.57 coverage leg.

[JarryShaw]

Verdict: GOOD TO GO at 62b96d993. Cross-reviewed on Sonnet; the author ran Opus.

• Scope: an independent AST scan of main finds the same 12 classes and 66 tests. The verifier's 13/72 included SchemaMetaReservedClassKwargsTests, which has no tearDown purge. After the change, the scan finds 0 such classes. ConstEnumRegisterFallbackTests is untouched.
• Order independence: each multi-test class passes 7 orderings, under both plain unittest and pytest (84/84 and 77/77). The two restore=True classes still leave 0 pcapkit modules behind. A Schema subclasses share one ABCMeta cache on Python <=3.10, so isinstance/issubclass return whichever answer was asked first #439 revert mutant fails the same 5 of 7 schema tests on main and on the branch.
• CI legs unchanged: protocols --exclude protocols/internet runs 1,231 tests and utilities runs 312, with identical per-test outcomes to main. Executed coverage lines are identical (27,861).
• Saving: the 7 modules in one pytest process went from 64–67 s to about 13 s.

Not blocking:

• Guard coverage: the guard catches the literal pattern that occurred, and addCleanup forms. It misses prefixes= keywords, aliases and mixin setUps.
• Order dependence: the dependence this PR isolates is a registry write-back that goes stale after a reload, contradicting its own docstring. I reproduced it and filed it as fix(protocols): _lookup_next_layer memoises the resolved class on a hit, so a reload leaves it stale #1559.

[github-actions]

Coverage: 90.02% (unit tier, Python 3.14, 62b96d993, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	4	93.55%
pcapkit/const	18795	842	2356	756	92.09%
pcapkit/corekit	2204	67	712	34	96.26%
pcapkit/dumpkit	258	2	90	2	98.85%
pcapkit/foundation	2797	132	986	48	94.18%
pcapkit/interface	108	7	36	5	91.67%
pcapkit/protocols	16992	229	4546	190	98.03%
pcapkit/toolkit	635	98	198	5	83.31%
pcapkit/utilities	431	4	124	4	98.56%
pcapkit/vendor	4409	2342	1006	157	43.25%

Per-file detail: the coverage-html artifact of this run.