perf(const): load the TCP and UDP app-type registries on first use (#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 (run one directory at a time, not as one make test)

• 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 (P7). pcapkit/const/reg/apptype/__init__.py now imports tcp/udp on first use: a PEP 562 __getattr__ for TCP/UDP/tcp/udp, and AppType.__registries__ holds a ModuleDescriptor per lazy registry until a read resolves it. Nothing on any import path needed them eagerly (import pcapkit, pcapkit.all, from pcapkit import *, extraction, engines, registry, transport/schema modules all checked). The other proposed importers (registry/protocols.py, fields/numbers.py, corekit/enum.py, const/reg/__init__.py) only import the base AppType, so they are unchanged.

Measured (3 runs each): -X importtime -c 'import pcapkit' 695/700/722 → 516/521/503 ms (tcp+udp were 206–214 ms, 29.5%, not 33%). tests/utilities/test_logging.py + test_const_enum_no_mint.py + test_const_apptype_split_unit.py: 43.0/42.8/43.1 → 37.7/37.4/40.8 s. One run of tests/const: 84.8 → 75.3 s, executing tcp.py 77 → 27 times (udp.py 77 → 26). Transport + application + capture round trip: tcp.py 76 → 16, udp.py 76 → 6 executions; 174.9/165.8/166.1 → 183.6/156.9/158.1 s (the first pair ran at load ~6, the rest at ~3). Estimate: ≤ ~0.19 s per cold re-import, minus ~0.1 s back for each registry a class then touches; per leg unmeasured.

Unchanged, pinned by tests: dir(), __all__, star-import, __registries__ keys/order/len/in and every read view, pickle/copy of members, a port registered before vs after first use, vendor output (regenerated all four from one IANA CSV under both inits: identical, and identical to the committed files). Suites green one directory at a time: const, corekit, protocols (all), foundation, interface, dumpkit, utilities, vendor, cli, toolkit, project, integration, top-level.

Contract changes for the breaking call:

• A bare import pcapkit no longer leaves pcapkit.const.reg.apptype.tcp/udp in sys.modules, nor TCP/UDP in the package __dict__.
• AppType.__registries__ is a dict subclass; __registries__.get costs 129 ns vs 56 ns, AppType.get 761 vs 691 ns.
• Test harness: tests/conftest.py now imports both registries before pinning its module snapshot. Without that, the restore dropped them after the first test that loaded them, and a later direct import built a second TCP/UDP; under xdist this failed test_the_audit_population_figures_are_the_measured_ones with 127 != 129 on 3.10–3.13. SharedImportProbe/PinnedSnapshotTests fail without the warm-up.

[github-actions]

Coverage: 90.03% (unit tier, Python 3.14, ed83fd807, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	4	93.55%
pcapkit/const	18858	842	2366	756	92.11%
pcapkit/corekit	2240	67	722	34	96.32%
pcapkit/dumpkit	258	2	90	2	98.85%
pcapkit/foundation	2766	129	980	47	94.23%
pcapkit/interface	108	7	36	5	91.67%
pcapkit/protocols	16996	230	4548	191	98.02%
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.

[JarryShaw]

Verdict: GOOD TO GO at ed83fd807. Cross-reviewed on Sonnet; the author ran Opus.

• Laziness: apptype.tcp and apptype.udp stay unimported after import pcapkit, pcapkit.all, a star-import and 12 other entry points. All 21 first-read routes resolve on first use: get, __registries__[...], iteration, ==, from ... import TCP, and so on. I checked the cold-import case myself.
• Behaviour: dir(), __all__, key order, pickle, copy and get identity all match main. The one visible change is that type(AppType.__registries__) is dict is now False. In 16-thread barrier runs, tcp.py and udp.py were each executed at most once. I don't think this needs breaking.
• CI fix: the conftest warm-up fixes the +2 counts failure. The old head fails 3 tests on the ordering repro; the new head passes 1039 of 1039 across tests/const, transport and project, in both orders. The laziness tests use child processes, so the warm-up can't hide a regression.
• Fail-before: 11 of 14 new tests fail on main's library. The 3 that pass are conftest guards, as intended.
• Numbers: import time goes from 695–722 ms to 499–521 ms. The two tests/const figures in the body didn't reproduce. A single run of 84.8 → 75.3 s came out at about 0–5 s net (89–95 s → 89–90 s), because the new test file costs about as much as it saves. AppType.get measured +15% rather than +10% (620 → 715 ns).

Non-blocking nit: the comment above __registries__ still says it is populated "once all four have been imported". That text is at pcapkit/const/reg/apptype/apptype.py:2400 and in the vendor template pcapkit/vendor/reg/apptype/apptype.py:355, and it is stale now that two entries start as descriptors.