fix(const,vendor): bring the AppType family onto EnumRegistry, stop it minting (#860)

[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

Step 2 of #860 (PR 2 of 2; the 8 non-AppType registries were #869). Brings AppType and its four transport subclasses onto EnumRegistry and stops both of its mint sites: _missing_'s 766 range branches, and get()'s own second, independent mint (PORT_{port}_{transport} -> 'unknown'). Both now build via a new _unregistered_member override reconstructing svc/port/proto (attributes the base's generic helper does not know about), per the owner's ruling: "only IANA registered ones are legit values ... get will not have sufficient information to create new ones."

All 766+1 branches convert uniformly, including the 8 branches that name a real (if IANA-assigned to a whole port span rather than declared individually) service, 7 distinct names with x11 appearing twice — x11, active-net, satvid-datalnk, vrml-multi-use, ircu, swx, flex-lm. Unlike FEATCode's earlier fix, there is no import-time self-mutation defect here to address by declaring members statically, so these are deliberately left as unregistered lookups too. The {TABLE}/{MISS}/{FLAG} data the generator derives from IANA's live CSV was not re-verified against a fresh crawl for this description; what I confirmed directly is that the class body the generator emits is byte-identical to what is committed (the earlier "557-line" figure was not reconstructible from the tree and is withdrawn).

AppType gains a working register() — previously absent; the base's generic one would have built a member with svc='<null>' once this class mixed in EnumRegistry without an override, since it calls __new__ with only one positional argument. Scoped to one per-transport registry, like register_alias, and now shares a _sanitize_identifier helper with it.

breaking, and this is why. TransportProtocol moves from power-of-two values (undefined=0, tcp=1, udp=2, sctp=4, dccp=8) to sequential auto() (undefined=0, tcp=1, udp=2, sctp=3, dccp=4), per the owner's follow-up ruling once |-joined values stopped being parsed at all; the stale comment claiming the values must stay power-of-two is deleted. AppType.get's proto parameter is documented public input (proto: 'TransportProtocol | str | int'), and _dispatch treats whatever int it is given as a whole rather than decoding it, on either numbering — so the renumbering changes what specific integers mean, not only what hand-composed ones do. Measured on AppType.get(80, proto=...), main vs. this branch:

proto	on main	on this branch
TransportProtocol.tcp | TransportProtocol.udp (int 3)	ValueError: 3 names no transport protocol registry	SCTP.http
bare 3	same ValueError	SCTP.http (identical to the composed case — _dispatch cannot tell them apart)
bare 4	SCTP.http	DCCP.unknown — silently, no exception, since a real member sits at 4 under the new numbering too, just a different one
bare 8	DCCP.PORT_80_dccp (minted under the old contract)	ValueError: 8 names no transport protocol registry

The 4 row is the one that matters most: a caller holding an integer written against the old spacing gets routed to a different, real transport registry with no error at all, not merely refused where it used to resolve. This consequence is accepted (the owner's ruling explicitly retires the power-of-two spacing) and deliberately left untested for the composed case specifically — "No need to add test around that honestly. This is an obsoleted path from a breaking change" — but it is disclosed here rather than only in a code comment, since it reaches further than composed values alone.

corekit/enum.py's own docstring described AppType's EnumRegistry mixin as future work ("stays as it is until tier two lands"); updated to reflect that it has landed, along with its get() docstring's str-valued-registry census (119 -> 124). tests/vendor/test_vendor_reg_apptype_generator_unit.py's own #770 pin is updated too: it asserted a bare-literal-vs-auto() distinction for undefined that no longer exists now that every TransportProtocol member is auto()-valued and cast.

All 5 crawlers (apptype, tcp, udp, sctp, dccp) regenerate byte-identically on a second run; git status --porcelain pcapkit/const is clean relative to this commit. Fixed 9 existing tests whose executable bodies changed — 8 that pinned the old minting/power-of-two behaviour directly, plus the #770 pin above (corrected from an earlier "6"; counted by comparing each test method's AST body against main, docstrings stripped) — and added 18 new tests in test_const_enum_no_mint.py against 1 removed, net +17 (plus prose/count updates there and in test_const_registry_protocol.py). 403 test methods pass across the complete tests/const/ and tests/protocols/transport/ directories; a further 87 across tests/vendor/ (excluding the live-network crawler-reachability test) and 55 across the TransportProtocol-consuming suites outside those directories (tests/corekit/test_fields_numbers_port_option_no_mint_unit.py, tests/dumpkit/test_nameless_enum_rendering_unit.py, tests/dumpkit/test_common_unit.py, tests/foundation/registry/test_protocols.py, tests/utilities/test_compat.py).

[JarryShaw]

NEEDS CHANGES on fc45d908a — cross-review (opus; author sonnet). The change itself is correct and unusually well covered; the head is red on a suite the PR never ran.

1. A test this PR invalidates was not updated, and it is the test for the generator this PR edits. Reproduced by me in the branch tree:

tests.vendor.test_vendor_reg_apptype_generator_unit   run=6 fail=1
FAIL: test_undefined_member_is_cast_rather_than_a_bare_literal
AssertionError: the cast is gone; expected "undefined = cast('TransportProtocol', 0)"

The file is not among the PR's 7 files (gh api .../files → 0 matches for it), so the ticked "make test passes" is false as written, and the "331 methods across tests/const/ and the touched tests/protocols/transport" claim is scoped to precisely the directories that exclude the breakage. Update that expected string and the method's docstring, which explains #770's fix in terms of a bare-0-versus-cast distinction that no longer exists in that shape; also re-check the sibling guard on the same method, re.search(r'\n[ \t]*undefined = 0[ \t]*\n', source), which is now vacuous. Its mypy twin test_undefined_member_infers_as_transport_protocol_under_mypy passes, so cast(..., auto()) still satisfies mypy — this is stale test text, not a broken fix.

2. One sentence missing from the description. The body says tcp | udp "now equals sctp numerically". The caller-visible form is sharper, and I measured it:

int(tcp|udp) = 3, and 3 is a member value
AppType.get(888, proto=tcp|udp) -> <SCTP.unknown: 888 [sctp]>
AppType.get(80,  proto=tcp|udp) -> <SCTP.http: 80 [sctp]>     # main raised ValueError

A composed proto now returns another transport's real declared member rather than raising. You accepted the breaking change and ruled the composition path obsolete, so this is not a blocker — but it belongs in the description in that form, because that is what a reader of the changelog needs.

Verified and not disputed, all re-derived rather than taken from the report: 766 + 1 mint sites converted with zero extend_enum left in _missing_ or get (classification exact — 754 unassigned, 4 reserved, and 8 real labels x11×2, active-net, satvid-datalnk, vrml-multi-use, ircu, swx, flex-lm); the sweep count is 1012 on an assertGreaterEqual floor, and the checker's own fixture table still flags value-derived names while exempting plain IANA labels; the _unregistered_member override reproduces __new__'s temp verbatim and leaves __registry__/__members__ untouched; register refuses a duplicate port, a colliding identifier and a non-identifier label, and register_alias/register_aliases still work after the _sanitize_identifier refactor; no persisted TransportProtocol value exists anywhere (zero hits across *.json/*.plist/*.txt/*.xml), no bitwise consumer in pcapkit/, and both flag() call sites pass a single element; the const file's entire 557-line class body is byte-identical to what the committed generator emits; and the three load-bearing tests fail on a real revert with the exact expected deltas (6147 != 6148, and the old sctp=4, dccp=8 map). 295 test methods measured green across tests/const/ and tests/protocols/transport/.

Left UNVERIFIED and not claimed as passing: test_const_enum_get and test_const_enum_lookup (budget), and the {TABLE}/{MISS}/{FLAG} data the generator derives from the live IANA CSV — that needs a network regeneration, so either do it or say "not checked" in the description.

Nit, not blocking: test_a_bare_int_composite_is_refused_as_a_whole was rewritten, not removed (your report said removed). The rewrite keeps every refusal assertion and switches the illustrative composite to tcp|dccp (== 5, provably not a member value); the 4 now-colliding composites are unavoidably no longer asserted refused. Its assertEqual(len(composites), 4) pins the collision by arithmetic complement, which brushes against the owner's "no composition test" ruling, and its comment about "lowest two bits" misses that tcp|sctp also collides.

[JarryShaw]

Addendum, narrowing my own verdict: the two suites I reported as UNVERIFIED are green on this head.

six tests/const/ suites together — no_mint, apptype_split, builtin_parity,
registry_protocol, enum_get, enum_lookup
Ran 231 tests in 175.800s → OK        run=231 failures=0 errors=0 skipped=0

So test_const_enum_get and test_const_enum_lookup add 21 methods over the 210 I first reported, and the measured green total on fc45d908a is 316 (231 const + 85 protocol). That removes one of the two gaps in my verdict.

The verdict itself does not change. tests/vendor/test_vendor_reg_apptype_generator_unit.py::AppTypeGeneratorShapeTests::test_undefined_member_is_cast_rather_than_a_bare_literal still fails on this head, and the file is still not in the diff. The one remaining genuinely unverified area is the generator's CSV-derived {MISS}/{TABLE}/{FLAG} content, which needs a network regeneration — so either run one or say "not checked" in the description.

[JarryShaw]

NEEDS CHANGES on 2715d1585 — cross-review (opus; author sonnet). The round-1 red test is genuinely fixed and its guards are not vacuous: stripping the wrapper (undefined = auto()) fails with the cast is gone; expected "undefined = cast('TransportProtocol', auto())", and appending a second bare undefined = 0 fails the regex loop. No test is red. What blocks is prose, and one undisclosed behaviour change I re-measured myself against main at 16879c36e:

proto	main	this branch
TP.tcp | TP.udp (3)	ValueError: 3 names no transport protocol registry	SCTP.http = 80
3	ValueError	SCTP.http = 80
4	SCTP.http = 80	DCCP.unknown = 80 — silent, no exception
8	DCCP.PORT_80_dccp = 80	ValueError

proto=4 is the one that matters: get()'s signature is proto: 'TransportProtocol | str | int', so a plain int is documented public input, and it now returns a different registry with no error. That is not "visible only to a caller composing bits by hand".

1. Two false statements in production comments — pcapkit/const/reg/apptype/apptype.py:49-53 and its generator mirror pcapkit/vendor/reg/apptype/apptype.py:143-146. "nothing that reads a real member observes this change" is refuted by proto=4 above, and "which AppType._dispatch already refuses as a whole regardless of numbering" is refuted by tcp|udp → SCTP.http: _dispatch does not refuse it, it resolves it, because 3 is now sctp. Your own quoted ruling ("treat it as a whole, instead of splitting") is satisfied — the word "refuses" is what is wrong.

2. Disclose the single-int changes in the PR body, not just the composed ones. breaking stays for exactly this reason.

3. tests/vendor/test_vendor_reg_apptype_generator_unit.py:232-238 says checking the literal "would silently pass a regression that dropped only the wrapper" — measured false, check 1 catches it. And the docstring's "every member — undefined included — has to stay wrapped, or the four [assignment] errors come back" is disproved: unwrapping tcp leaves the mypy test passing, SKIPPED = 0. Only undefined is guarded, which is all #770 needed.

Confirmed: 87 / 251 / 152 / 55 methods reproduce exactly (the 7 transport errors are the missing generated examples/captures/ fixtures, not this PR); the no-mint floor is met at 1012 with zero slack, apptype.py contributing 768; extend_enum survives in the const file only at :2697/:2781 (register/register_alias). Byte-identity of the regenerated table rests on a crawl neither of us re-ran.

[JarryShaw]

NEEDS CHANGES on 01ae462ed — cross-review (opus; author sonnet). The corrected comment is byte-identical in both copies (sha256 6cba08c9…, 2186 bytes, no braces so no {{/}} escaping in play) and every claim in it checks out: _dispatch is literally the same method on this branch and on main (code + docstring identical with comments stripped), the lookup is a plain dict.get at apptype.py:2475 on a plain dict at :2315, and a real member sits at 4 on both trees. But the new docstring replaces two false statements with another one.

1. "only undefined's own wrapper is load-bearing here" is false. I measured all three shapes with the test's own mypy invocation:

baseline (all wrapped)            status=0  assignment=0  Success: no issues found in 1 source file
undefined = auto()  (unwrapped)   status=0  assignment=0  Success: no issues found in 1 source file
undefined = 0       (#770 shape)  status=1  assignment=3  Found 765 errors in 1 file

Unwrapping undefined alone is also clean, so its cast is not load-bearing for the mypy companion either. auto() reveals as Any, and Any is assignable to TransportProtocol — which also makes the stated justification self-defeating, since Any is precisely what does not error; #770's literal errored because 0 infers as int. The honest statement: no member's cast is load-bearing under auto(); what the mypy test guards is a revert to a bare int literal, and what the source-text test guards is the exact wrapped string.

2. Same paragraph: "#770's own four [assignment] errors" measures as 3, not 4 — :2298, :2322, :2793, plus 762 [arg-type]. The fourth site stopped erroring because proto is now annotated 'TransportProtocol | str | int', which accepts an int default.

**3. The "since it's no longer a Flag, | joined values …"quote is attributed to #860 and is not there.** It is in **#836** — 2 comments in #836 matchno longer a Flag, 0 in #860. _dispatch`'s own comment in the same file cites #836 correctly, so one file attributes one quote to two places. Pre-existing, but the block is open.

Two errors in the body were mine and I have fixed them. "Fixed 6 existing tests" was understated: comparing each test method's AST body against main with docstrings stripped gives 9 changed, the #770 pin among them. And "added 17 new tests" is 18 added against 1 removed, net +17. I also withdrew the "557-line class body" figure — it is not reconstructible from the tree (the template spans :216–:813, giving 598/597/513/512/382/381 by six natural readings, none of them 557).

Optional, and I agree with your reasoning for leaving assertEqual(len(composites), 4) out: 0 < len(composites) < len(candidates) pins "the filter is neither a no-op nor total" without naming 4 or which values collide, and catches a renumbering under which nothing collides — which assertGreater(…, 0) cannot. Your call.

Byte-identity across a live crawl remains unreproduced by either of us, and the body says so.

[JarryShaw]

NEEDS CHANGES on 7397976b9 — cross-review (opus; author sonnet). All three round-3 defects are fixed and every number now checks out, but two claims in the newly written prose are wrong, and one of them is a fresh misattribution inside the paragraph whose job was to fix a misattribution.

1. "since #860" on get/get_all's proto union is the wrong issue (tests/vendor/test_vendor_reg_apptype_generator_unit.py:299-302). 'TransportProtocol | str | int' was introduced by #836, and it is already present in this commit's own parent:

$ git log --oneline -S "proto: 'TransportProtocol | str | int' = TransportProtocol.undefined" -- pcapkit/const/reg/apptype/apptype.py
14d3d3dc7 fix(reg)!: drop TransportProtocol's IntFlag base now that nothing composes (#836)
$ git show 60b85e3a4:pcapkit/const/reg/apptype/apptype.py | grep -c 'TransportProtocol | str | int'   → 3

So the 4 → 3 story has the mechanism right and the provenance wrong: #836 caused the 4 → 2 drop; #860's only contribution is adding _unregistered_member, taking it 2 → 3.

2. pcapkit/const/reg/apptype/apptype.py:2490-2492 (vendor :445-447) names a value as refused that resolves. That block sits after the __registries__.get hit-return, so it is the miss path, and it opens "everything that reaches here names no registry" — then says a bare int "is refused exactly the same way whether it is a single stray bit, e.g. 17, or a composite of several real transports, e.g. 3 (tcp | udp)". Measured: 17 raises, 3 returns SCTP and never reaches that block. It contradicts lines 56-58 of the same file, which this round wrote. Three softer instances carry the same staleness — :2444-2446, :2527, :2545-2547 and their vendor twins all assert a hand-built composite is "refused".

Cosmetic, worth folding in: #836 is a PR, not an issue (is_pr: true; #860 is_pr: false). Three of the new paragraphs say "GitHub issue #836" while sibling comments in the same files say "this PR (#836)".

Confirmed, and not to be redone: the mypy numbers are exact — undefined = 0 gives 3 [assignment] (:2303, :2327, :2798) plus 762 [arg-type] for 765, and the 762 is literal, since 766 _unregistered_member calls sit in _missing_ and the 4 that pass a real member (x11 ×2, ircu, reserved) do not error. reveal_type gives Any unwrapped, TransportProtocol wrapped, int for the literal. _unregistered_member really is new (0 occurrences at 60b85e3a4). Both comment copies are byte-identical (sha256 1160f8c5f0…, 2518 bytes). The #860 auto() quote is verbatim in #860, and the Flag quote is now cited as #836 in all four files. The composites bounds are each non-vacuous — if True trips 8 not less than 8, if False trips 0 not greater than 0. Token comparison confirms the only executable delta is self.assertLess(len(composites), len(candidates)): the const module is token-identical including every string literal. 6 and 216 methods, 0 failures, 0 skipped — so the mypy test ran rather than skipping into a false green.

[JarryShaw]

Recorded, verbatim:

undefined direct uses 0.

then other real transport use auto. so we don't have to define a start and the undefined declaration is explicit

So the GOOD TO GO on 934f26f60 is superseded — label goes to review: needs-changes and round 7 is in flight. I measured the proposed shape before briefing it, and it is a pure simplification: drop _start_ = 0, write undefined = cast('TransportProtocol', 0), leave the four real transports on auto().

NUMBERING {'undefined': 0, 'tcp': 1, 'udp': 2, 'sctp': 3, 'dccp': 4}
  proto=3 -> http [80 - sctp]      proto=4 -> unknown [80 - dccp]      proto=8 -> ValueError
mypy status 0 | Success: no issues found in 1 source file

Identical numbering, identical dispatch, and mypy stays clean — because auto() after an explicit 0 continues from it, so _start_ was only ever needed to make undefined itself an auto(). One thing worth noting for the history: this puts the #770 pin's expected string back to cast('TransportProtocol', 0), which is what it asserted before round 2 changed it — the wrapper is what keeps a bare literal from inferring as int, which was #770's original point, and it is now load-bearing again for exactly one member rather than for none.

Everything else in the 934f26f60 verdict stands; the round-7 diff should be the two declaration lines plus the prose and the pin that describe them.

[JarryShaw]

NEEDS CHANGES on e51a0c6b1 — cross-review (opus; author sonnet). The declaration change is right and everything measurable checks out. What broke is the comment this round exists to reword: the edit dropped a clause, and I verified all three consequences myself.

Round 6 read ... #808 already dropped that base once / nothing built a composite, and GitHub PR #836's ruling later / retired ``|``-composite decoding entirely: "since it's no longer a Flag ...". Round 7 collapsed that to:

# GitHub issue #808 already dropped the ``IntFlag`` base once
# retired ``|``-composite decoding entirely: "since it's no longer a

So: "once retired" has no subject; the Flag quote is now attributed to #808 rather than to #836's ruling, contradicting the settled account that #808 is the issue and #836 the PR that implemented it; and the back-reference at :63 — "exactly what #836's ruling above asks for" — now dangles, because grep -n "#836" puts its first occurrence at :63, with the next at :102, nothing above. Same in the vendor twin at :143-144 and :156; the two are currently byte-identical (sha256 1f6ad8bd5c20…, 42 lines each), so fix them together via the template. Restoring the dropped clause verbatim fixes all three at once.

The _start_ sweep missed a fifth passage of the kind it fixed four of: tests/const/test_const_apptype_split_unit.py:1352 still says "the five now number sequentially from 0 via auto() instead". Only four are auto()-valued; undefined is a direct 0. Pre-existing, but it means "that is all of them" is not true.

Confirmed clean and not to be re-derived: numbering {undefined 0, tcp 1, udp 2, sctp 3, dccp 4} with _start_ = 0 gone and dispatch unchanged across proto as member, as 3, 4, 8, 17, 0. The inverted mypy story holds in all three shapes — head clean; undefined unwrapped → 765 total with 3 [assignment] at :2308/:2332/:2819; tcp unwrapped → clean — and reveal_type shows exactly why: int for the bare literal, Any for a bare auto(). The #770 pin discriminates on every reachable mutation, and a second in-class undefined = 0 is rejected by aenum itself (attempt to reuse name) before the regex is needed. Reverting to cast(..., auto()) without _start_ also trips test_transport_protocol_undefined_still_zero with 1 != 0 — a real second line of defence. Rebase verified: exactly 1 commit on 687721091, zero file overlap with it, and _dest_path appears in none of the PR's 8 files, so main's new @classmethod strands nothing. 216 methods reproduce with 0 skipped, so the mypy-gated test really ran.

One caveat worth recording: '_start_' in vars(TransportProtocol) reads True even now — that is aenum's own default (TP._start_ == 1, same as a never-declared control class), not a survival of the deleted line. MERGEABLE/BLOCKED is a required context not yet reported on the new head; every reported context is SUCCESS except Gate (full suite, Python 3.14), which is SKIPPED.

[JarryShaw]

GOOD TO GO on f58fb759b. Re-granted by diff rather than by re-review, since round 8 is exactly the two items I raised and nothing else.

git diff e51a0c6b1 f58fb759b is three files, +7/−2: the dropped clause nothing built a composite, and GitHub PR #836's ruling later restored in both copies, and the fifth over-generalisation at tests/const/test_const_apptype_split_unit.py:1351-1353 rewritten to distinguish undefined's direct 0 from the four auto() members.

All three consequences are discharged. #836 now first appears at :51, so the back-reference at :64 — "exactly what #836's ruling above asks for" — resolves two lines above itself instead of dangling; "once" has its subject again; and the Flag quote is attributed to #836's ruling rather than to #808. I re-derived the byte-identity myself rather than taking the report: the 44-line comment block hashes identically in both files (sha256 a1f4ef4be37e…), up from the 42-line 1f6ad8bd5c20… baseline by exactly the two restored lines.

Still one commit on 687721091, still 8 files. Everything else in the e51a0c6b1 verdict carries over unchanged — no executable line moved this round, so the numbering, the dispatch table, the inverted mypy measurements, the #770 pin's discrimination and the 216-method run all stand as measured there.

The author also named the files behind its "55 methods" figure, which the previous review could not verify: tests/corekit/test_fields_numbers_port_option_no_mint_unit.py (16), tests/dumpkit/test_nameless_enum_rendering_unit.py (6), tests/dumpkit/test_common_unit.py (14), tests/foundation/registry/test_protocols.py (14), tests/utilities/test_compat.py (5) — and put the list in the commit message, so it no longer depends on a hand-back report.