fix(const,vendor): stop EtherType and Socket minting on an unrecognised value

[JarryShaw]

Please follow the guide below

• 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, 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

Description of your pull request and other information

Closes the remaining scope of #775. EtherType (52 branches) and Socket (1) were the last two registries whose _missing_ still minted a permanent member for an unrecognised value; both now return a non-registering cls._unregistered_member(value, name). Measured on this head with the same AST walk the issue tracked, counting extend_enum inside _missing_/get across every class in pcapkit/const/:

before (edc1b32e0):  52  reg/ethertype.py  +  1  ipx/socket.py  +  1  mh/cga_type.py  = 54
after  (f210ddc71):   1  mh/cga_type.py

That single remaining site is the standing CGAType carve-out — its mint is not an IANA-style range — so this takes the issue from its original 1,169 sites to the one deliberate exception.

Behaviour, measured both directions with the tree under test asserted and printed each run. Before, EtherType(0x0888) grew __members__ 160 → 161 and a repeat lookup returned the identical object; after, it returns <EtherType.Xyplex_0x0888: 2184> with __members__ unchanged, absent from _value2member_map_, and equal-but-not-identical across lookups. Same shape for Socket(0x0010).

Two things reviewers should look at rather than skim.

pcapkit/vendor/ipx/socket.py's output was byte-regenerated offline and diffed identical, because Socket.LINK is None so its crawler never touches the network. pcapkit/vendor/reg/ethertype.py's was not — that crawler fetches the live IANA CSV, which is off-limits here. Its correctness rests on a mechanical substitution verified to match exactly 52 lines, plus running the generator against the already-committed ETHERTYPE_862_FIXTURE_CSV and checking the emitted lines. That is weaker than byte-identity and is stated as such.

Preserving each branch's existing hex-suffixed name — required, since this change is about not registering rather than renaming — reintroduces the value-derived name shape that test_const_enum_no_mint.py's is_manufactured() sweep forbids. Rather than weakening that predicate, the PR adds a narrow, documented HEX_SUFFIXED_NAME_EXEMPT_PATHS covering exactly these two files, on the grounds that the collision the check guards against cannot occur once _unregistered_member never registers. That is a guard being narrowed, so it deserves the sharpest look in this diff.

pcapkit/corekit/enum.py is touched for one docstring claim only — it said the mint was "kept minting, on three registries (EtherType, Socket, CGAType)", which this change makes false; it now names CGAType alone. No logic change. Three of the four test files beyond the two required ones carry pre-existing assertions that this change genuinely falsifies. Correction: tests/vendor/test_ipx_socket_unit.py is not one of them — its code tokens are identical to edc1b32e0 (1129 vs 1129) and the edit is comment-only; its EXPECTED_MISSING_NAMES still carries 0x0010: 'Registered by Xerox_0x0010' and still passes, because the name was preserved.

One commit on edc1b32e0. tests/const/test_const_enum_lookup.py 13/0/0 and tests/vendor/test_vendor_missing_body_unit.py 6/0/0, with test_const_enum_no_mint.py 69/0/0, test_const_ethertype_862_unit.py 9/0/0, test_const_registry_protocol.py 74/0/0 and tests/vendor/test_ipx_socket_unit.py 10/0/0 green after the edits. EXPECTED_CRAWLERS_WITH_CONSTANT_MISS_LINES stays at 35 — both crawlers still build miss by append, so they remain invisible to that static sweep, and the pins added here are behavioural instead.

[JarryShaw]

GOOD TO GO on f210ddc71 — cross-review (opus; author sonnet). Two cosmetic findings, neither blocking, both verified by me.

The headline re-derived by a different route than the author's (per-class rather than per-file): 54 → 1 extend_enum calls inside _missing_/get across pcapkit/const/, sole survivor the documented CGAType carve-out, which this diff does not touch.

The conversion was checked exhaustively, not sampled. Every range branch of both _missing_ bodies was AST-extracted and probed at lo, hi, mid, lo−1, hi+1 — 257 probes, 177 live. On this head: 0 probes grew the registry, 0 landed in a lookup table, 0 returned an identical object on a second lookup, 0 name or int() mismatches. On edc1b32e0 the same probes grew it 152 times, each members+1 v2m+1 via extend_enum. The 53 probes that raise are identical between trees, so no branch changed reachability — that is the boundary check, not a mid-range spot check.

The narrowed guard is justified but shape-blind, and I reproduced that myself. is_manufactured() exists to catch a name built from the value, because such a label risked a __members__ collision between minted members; _unregistered_member is three statements touching neither _member_map_ nor _value2member_map_, so nothing it is handed can collide — the harmlessness follows from the method, not the name. Measured: all 53 exempted call sites share one identical suffix expression, and 0 manufactured names outside the two files are exempted, so the predicate still has teeth everywhere else. But the exemption is keyed on path, so within those two files any other manufactured shape also passes:

mutated line 540 to: return cls._unregistered_member(value, 'Xyplex_%d' % value)
UnregisteredMemberNameIsBareTests -> run=1 failures=0 errors=0

Non-blocking, because the widened area is harmless for the same structural reason the intended 53 are. Worth a follow-up to key the exemption on the suffix expression rather than the path.

One claim in the PR description is wrong and I am correcting it as the author of that text. I wrote that the four extra test files were changed because pre-existing assertions became false. That holds for three of them, but tests/vendor/test_ipx_socket_unit.py is code-token identical — base 1129, head 1129, identical=True; the edit is comment-only. Its EXPECTED_MISSING_NAMES still carries 0x0010: 'Registered by Xerox_0x0010' and still passes, precisely because the name was preserved.

Also confirmed: pcapkit/corekit/enum.py is docstring-only — AST with docstrings blanked and the token stream with strings collapsed are both identical (543 vs 543 tokens), comments identical 25 vs 25. The other three test files removed no coverage — each removed assertion is replaced by its exact negation, call_count >= 1012 was strengthened to >= 1065, and a new method sweeps EtherType where it was previously skipped. All six suites reproduce exactly: 13/6/69/9/74/10, 0 failures, 0 errors. EXPECTED_CRAWLERS_WITH_CONSTANT_MISS_LINES = 35 is byte-identical and correct — socket.py assigns an empty ast.List and ethertype.py a ListComp, so neither enters that static sweep. vendor/ipx/socket.py's output is byte-identical to the committed const file when process() is run offline (15 of 15 interior lines), and for ethertype.py every one of the 56 committed branch lines matches one of the template's two emitted shapes byte-for-byte, with the (lo,hi) and name sequences identical to base.

Unverified and stated as such: whether a live IANA crawl today would still yield these same 52 rows. No crawl was run, and that proposition was equally unverified at edc1b32e0.