test: re-import pcapkit only where isolation needs it (#1065)

[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

• 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?

• 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
• chore — anything else

---

Description of your pull request and other information

Closes #1065. Closes #1062.

• 204 setUp purges become tests._support.reimport_once_per_class(self). Each class gets one private pcapkit import, which its own tests share. A setUpClass purge would not work, because the conftest restore swaps the shared session import back in before every test. Plain unittest still leaves the class's import behind, so test: three modules pass alone but fail together, and pytest reports it green #981's reproduction (test_http_unit then test_base_class_contract on 32bcfba15^) still gives 1 failure and 3 errors.
• Fixed state that tests leaked to later tests in the same class or module: pcapng option tables, the zero-pad ledger, leftover schema subclasses, an uncollected class, Extractor registries, a scapy-less import. ConstEnumRegisterFallbackTests and TCPHTTPDispatchTests keep a per-test purge.
• tests: audit the suite for redundant cases that add no unique coverage #1062 trims: removed two duplicate tests, the pcapng sweep now runs once instead of four times, and the AppType sweep is sampled (37,173 to 882 subtests, the same 4 arc sets). The const: make EnumRegistry.get's NO_DEFAULT a sentinel object, not -1 #857 -1 test now discriminates again.
• Times, one module per process under pytest: the 129 affected modules took 2,156 s before and 483 s after. They also pass under plain unittest, in CI-style -n 4 directory groups, all unit modules in one process (forward), reversed in four chunks, and the PyPCAPFile engine leg on 3.10.

[JarryShaw]

NEEDS CHANGES on 1d986fd9f: CI shows the cross-test leakage the per-test purge used to hide. main is green on these same tests.

• Python 3.14 (job 112112341217): 4 failed, 2604 passed in 54 s, against about 16 min on main.
  • test_option_roundtrip_unit — the pcapng-option/opt_comment_1 round trip parses as PARSE, not OK.
  • test_protocol_code_registration_unit — the backward-compatible dispatch table holds a resolved IPv4 class, not a ModuleDescriptor.
  • test_protocol_base_unit — Raw is not found in (Raw, 'RAW'): two different Raw classes, one from each import of pcapkit.
  • interface/test_core — test_every_shipped_engine_has_a_constant sees an extra engine.
• Engines PyPCAPFile 3.10 (job 112112341590): 3 failures in test_pypcapfile_engine.py.
  • 'types.SimpleNamespace' object is not subscriptable.
  • ipv4_reassembly was called 0 times where the test expects once.
• All five Python legs (3.10–3.14) fail, so Required checks passed fails too.

These only show up when many modules share one process. The one-module-per-process runs in the PR description could not see them. The fix has to hold under xdist with the full unit selection.

[JarryShaw]

Cross-review verdict on 1d986fd9f: NEEDS CHANGES (ran on Sonnet; author Opus). This is in addition to the CI failures above.

• Root cause: under pytest the conftest fixture restores module bindings before every test, so the new setUpClass purge has no effect. Registry contents, __proto__ tables and engine tables are not restored. Any test that mutates them without cleanup now leaks into later tests in the same worker.
• Polluters bisected (candidate plus victim in one process):
  • foundation/test_extraction.py leaves an engine registered in Extractor.__engine__. That breaks interface/test_core. I reproduced this on a clean checkout.
  • 13 modules that dispatch through IPv4 resolve its lazy __proto__ entry. That breaks test_protocol_code_registration_unit.
  • test_dispatch_default_resolution_unit leads to the Raw mismatch between import generations.
  • PyPCAPFile on 3.10 is a stale package attribute seen by pre-3.12 mock target lookup.
• Two more order-dependences, which show up only when tests run in reverse:
  • foundation/registry/test_protocols.py leaves UNITPROTOCOL registered.
  • test_layer_placement_unit sees OSPF/RARP already resolved.
• Holds:
  • test: three modules pass alone but fail together, and pytest reports it green #981 still gives 1 failure and 3 errors on 32bcfba15^.
  • All 11 run_unittest_leg.py legs pass.
  • Five of the seven earlier fixes are legitimate hygiene. The test_extraction restore is partial, and "accepts either form" is incomplete because the same hazard recurs in other modules.
  • The -1 test discriminates again.

[github-actions]

Coverage: 88.74% (unit tier, Python 3.14, ae3ceb7ad, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	4	93.55%
pcapkit/const	18757	1035	2342	846	90.23%
pcapkit/corekit	1874	91	578	22	94.33%
pcapkit/dumpkit	136	0	40	0	100.00%
pcapkit/foundation	2422	143	842	34	92.62%
pcapkit/interface	112	7	40	5	92.11%
pcapkit/protocols	15653	187	3942	162	98.19%
pcapkit/toolkit	487	71	144	3	84.15%
pcapkit/utilities	429	4	122	4	98.55%
pcapkit/vendor	4409	2359	1006	158	42.84%

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

[JarryShaw]

Cross-review verdict on 2efb76dbb: NEEDS CHANGES (one small fix) (ran on Sonnet; author Opus)

• The one regression. In tests/foundation/registry/test_protocols.py, test_register_protocol_validates_and_updates_registry (:410) and the test near :461 register UNITPROTOCOL without _guard_registry.
  • Under the per-class import, test_register_protocol_stays_quiet_when_nothing_is_displaced then sees an overwrite warning.
  • I reproduced it on a clean checkout of this head: SUBFAILED(case='fresh-key').
  • The same order passes on main. CI's order happens to hide it.
  • Fix: self._guard_registry(registry.protocol_registry, 'UNITPROTOCOL') in both tests.
• Everything else holds:
  • test: three modules pass alone but fail together, and pytest reports it green #981 is still caught: 1 failure and 3 errors on 32bcfba15^, in 2.1 s instead of 72 s.
  • All 11 run_unittest_leg.py legs pass.
  • The 170 CI unit modules pass under xdist -n 4 --dist load, in 7 chunks. Last round's polluter/victim mix passes at -n 3 and -n 4.
  • 119 changed modules pass in reversed order, apart from the one above.
  • tests/const reversed: peak memory 412 MB in 89 s, against 488 MB in 271 s on main.
  • No assertion is weakened.

[JarryShaw]

Cross-review verdict on ae3ceb7ad: GOOD TO GO (ran on Sonnet; author Opus). Confirmed by this head's own CI: 72 checks pass, including Required checks passed.

• The UNITPROTOCOL leak is fixed. The only change since 2efb76dbb is two _guard_registry(registry.protocol_registry, 'UNITPROTOCOL') lines, which I confirmed with git diff. The previously failing pair now passes on a clean checkout.
• Reversed-order runs: each of the 129 changed modules passes run alone in reverse order, under both pytest and plain unittest (258 runs).
• Carried over from the 2efb76dbb review:
  • test: three modules pass alone but fail together, and pytest reports it green #981 is still caught: 1 failure and 3 errors on 32bcfba15^.
  • All 11 legs pass.
  • The 170 CI modules pass under xdist -n 4.
  • Peak memory is lower than main.
  • No assertion was weakened.
• Speed: the 129 changed modules take 483 s, against 2,156 s before. test_http_unit takes 1.07 s, against 77 s on main.