perf(reg): drop the per-member annotation and TransportProtocol.get() calls

[JarryShaw]

• 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 — bounded selections instead, counts below
• Added a changelog entry — N/A — changelog 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

Fixes #744. Fixes #768. Both are one-line changes to the same generator forcing a full regen of the same five files, so one PR avoids conflicting ~12,000-line diffs.

#744 drops : {self.NAME!r} from each member's assignment (apptype.py:717), plus the # mypy: disable-error-code=assignment pragma it forced. Removes 86,838 bytes. Autodoc's sphinx.util.typing.get_type_hints() re-resolves the whole class per member — 378.8s on CI, the docs build's largest stall; measured on this PR, deploy-pages drops 567s → 213s.

#768: flag() now emits TransportProtocol.tcp instead of TransportProtocol.get('tcp') — ~21x cheaper, 23,941 of 23,942 sites rewritten (the 23,942nd, in _dispatch, resolves a caller-supplied string and stays get()). 7 bytes/site, 167,587 bytes. Covers #764's 4 range-row conditions too.

Regen: one pinned IANA snapshot; unmodified crawler reproduces the five files byte-identically first, then the diff after both changes is mechanical throughout (254,615 bytes total: 190 pragma + 86,838 annotation + 167,587 .get()). Exhaustive member-set (name/port/svc/proto) compare across all 12,391 members: identical. #764's behaviours re-confirmed.

Tests: updated #760's regex for the new form and tightened its trailing match; added generator pins (0%→20% is the wrong framing — the module was already at 19% via test_vendor_dest_path_unit.py/test_crawler_reachability_unit.py; the new file's flag() pin takes it to 20%), each shown to fail without its fix. make isort clean.

mypy: removing the pragma surfaces 4 pre-existing errors (112→116), all pre-existing TransportProtocol.undefined-as-default mistyping in the base template, unrelated to either fix. Advisory-only in CI; left unfixed/unsuppressed rather than silently re-adding the pragma.

[JarryShaw]

Decision needed: what to do with 4 mypy errors this PR unmasks.

Removing the annotation makes # mypy: disable-error-code=assignment redundant, so the PR drops it from all five generated files (it was on 5, not the 4 I had guessed). That pragma was also hiding four errors unrelated to this change. Measured on the apptype file alone:

main (pragma present) :  Success: no issues found in 1 source file
this branch (removed) :  4 errors
  :2258  Incompatible types in assignment (expression has type "int", variable has type "TransportProtocol")
  :2282  Incompatible default for parameter "proto" (default has type "int", parameter has type "TransportProtocol")
  :2371  Incompatible default for parameter "proto" (default "int", parameter "TransportProtocol | str")
  :2429  Incompatible default for parameter "proto" (default "int", parameter "TransportProtocol | str")

Whole-package: 112 errors in 38 files → 116 in 39. All four are in the base template's static code (__new__, _dispatch, get, get_all), every one about TransportProtocol.undefined used as a default.

Root cause, isolated. undefined = 0 is a bare int literal so mypy infers its attribute type as int, while tcp/udp/sctp/dccp are auto()-valued and infer as Any — which is why only undefined trips it. mypy has no aenum plugin, so the class is treated as plain rather than as an enum. Pre-existing and independent of this PR: stripping only the pragma on unmodified main reproduces all four, and none of the ~12,391 per-member lines contributes a new error.

Three options:

1. Leave them unsuppressed and file them as their own issue. mypy is advisory in CI (continue-on-error: true in lint.yml), so nothing breaks, and ci(lint): the workflow's recorded analyser counts have drifted (mypy 115/39 vs 112/38) #753's whole premise is that recorded drift beats hidden drift. The count in lint.yml's header would need updating to 116/39 — or the issue would note it deliberately.
2. Re-add a narrow suppression naming only [assignment] on those four lines, rather than the file-wide pragma. Keeps the count at 112 and keeps the fix honest about what it is suppressing.
3. Fix the inference. Likely needs an aenum stub or an explicit annotation on the TransportProtocol members; I have not established that it is achievable without one, so this may not be a real option today.

I would take (1): they are real, they were always there, and this PR's job is not to carry them. Recommending it rather than deciding, since it moves a published count.

This does not block the PR — CI is fully green, including deploy-pages, which means the docs build no longer hangs.

[JarryShaw]

take (1) then.

[JarryShaw]

Ruling recorded: option (1) — leave the four errors unsuppressed. Filed as #770 with the root cause, the isolation showing they are pre-existing, and the four sites (__new__, _dispatch, get, get_all).

#770 also carries the one follow-up this creates: lint.yml's header pins mypy at 112 errors / 38 files measured at 932cb48d1, which #761 has just repinned and is merge-ready. That pin stays honest because it names its commit, but once this PR merges the live figure is 116/39 — so whichever of the two lands second leaves the other reading stale unless the count is refreshed. Flagged there rather than silently letting them drift apart.

[JarryShaw]

Cross-review at 9ae0e2991: GOOD TO GO (opus, a different model from the author).

All four prose corrections landed. The worked example is now character-for-character identical to tcp.py:946 (same sha256 once the indent is stripped), and the byte figures re-derive from the real member counts: 12,290 × 7 + 101 × 8 = 86,838 for the annotations — 7 bytes where the class name is 3 letters, 8 where it is 4 — plus 7 × 23,941 = 167,587 for the .get() rewrite, which is a flat 7 per site because TransportProtocol.get('sctp') is 29 → 22 as well. With the five pragmas that is 190 + 86,838 + 167,587 = 254,615, matching the measured file-size delta to the byte.

git diff --stat 1a4b77353..9ae0e2991 -- pcapkit/ is empty, so the generator and all five generated files are untouched this round and last round's byte-identical regeneration proof carries over. The assertIn → condition.endswith(claim) tightening was checked against a truth table over the real emitted conditions: it keeps every real shape, still rejects a wrong proto and a negated is not, and additionally rejects both directions of the prefix collision the NOTE describes. The generator emits the claim as the suffix by construction, so requiring it to be final is correct rather than over-tight.

One correction to the headline number, which I had overstated. The deploy-pages baseline is noisier than a single pair suggests:

main 57b2c1761  567 s     main 60bb62d8f  574 s     main 1a852698b  717 s
PR   1a4b77353  213 s     PR   9ae0e2991  146 s

Round 2 touched only tests/, which the docs build never reads, so 146 s versus 213 s is runner variance on identical docs input, not a further improvement. The reduction is large and unambiguous — roughly 3-5× — but it should be stated as a range, ~570-720 s on main against 146-213 s on this branch, rather than as a precise pair. I will not hold the PR for that.

Also worth recording: the commit message calls the pragma "now-redundant … on all five generated files". On four it is redundant; on the base module it was suppressing four unrelated errors, since keeping it there alone restores 112 errors in 38 files exactly. The decision to leave them visible is settled and tracked in #770.

Merge with current main 1a852698b re-confirmed conflict-free, with Ran 115 tests … OK on the merged tree across test_tier_guard (#755), test_isort_clean (#761) and both of this PR's test files. Unpublished and awaiting the owner.