[JarryShaw]

• Searched for similar pull requests
• Followed the coding style — I ran pylint, mypy and isort on protocol.py; they report nothing new in the changed lines
• make test passes, and a test case covers the change — I ran only the two unittest legs, tests/project and the touched modules (below)
• Added a changelog entry — N/A, added centrally

What is the purpose of your pull request?

• fix — corrects a defect

---

Closes #1308
Closes #1312

#1312. unpack() reused the made schema, so raw bytes items reached read() and crashed with AttributeError.

• When it re-parses: only if make() was given raw or Schema items, and the made schema still holds raw bytes where parsing yields a schema. The parsed header then replaces the made one.
• Malformed items: if the octets do not parse, it raises ProtocolError chained from the parse error. It catches only BaseError and struct.error.
• Which protocols: on main this crash hit 8 protocols: HIP, IPv4, TCP, HOPOPT, IPv6_Opts, MH (including a bare MH()), SCTP and IPv6_Route.
• Fuzz: 1,500 seeded random inputs per protocol produce 0 non-library exceptions.

#1308. _make_index_namespace locates the _make_index call with ast and resolves against that call's first argument. All 111 call sites pass a plain name. 12 of them, in hip.py, pass loop variables rather than parameters (group, suite, cipher, mode, reg, tf). When the first argument is not a parameter, or the source is unavailable, it falls back to main's all-arguments check. So x = b; self._make_index(x) behaves as on main.

Behaviour change: test_make_index_str_unit pinned the ambiguous-name raise, so I updated it.

Cost: best of 3, 10,000 constructions each:

	main (4849d16e4)	branch
TCP()	89.4 µs	93.0 µs
IPv4(protocol=253)	109.9 µs	116.0 µs

The host load average was about 6 during the runs.

Evidence:

• New tests: test_protocol_make_unit.py gives 25 failed, 4 passed against main's protocol.py, and 12 passed with the fix.
• Unit legs: protocols/internet 529 tests, 0 failures, 0 errors; the rest of protocols 918 tests, 0 failures, 0 errors.
• tests/project: 389 passed, 1 skipped.
• Fixtures: info is identical to main on all 23 fixtures.

[JarryShaw]

Cross-review verdict on 2a7ca1c49: NEEDS CHANGES (Sonnet reviewed; author Opus). There is one blocker, and I re-derived it myself.

1. The broad except in unpack() keeps #1312's crash for malformed raw items. When the re-parse fails, the made schema is kept with raw bytes in it, and read() then raises the original error. IPv4(options=[b'\x02', b'\x94\x04\x00\x00'], protocol=253) gives AttributeError: 'bytes' object has no attribute 'type' at ipv4.py:621. The ProtocolError: Field data resolved to a negative length that the except swallowed was the right answer. The reviewer's fuzz of 1,500 inputs per protocol ended in AttributeError for IPv4 497 times, HOPOPT 199, TCP 284, IPv6_Opts 164 and MH 27.

Fix: when the re-parse fails and raw items are present, raise a ProtocolError (chained from the parse error), and narrow the except to the library's parse errors.

2. Not blocking. _holds_raw() walks the whole schema on every made construction, which adds about 11% even with no raw items: TCP with no options goes from 93–96 µs to 105–109 µs. Short-circuit it, for example with a flag set only when raw items are supplied. Separately, a first argument that is not a parameter (x = b; self._make_index(x)) now resolves nothing, where main resolved it. All 12 such sites in hip.py pass namespace=, so nothing in-tree breaks, but the PR body's "all 111 pass a plain name" should say that 12 of them are not parameters.

Held up:

• fix(protocol): a raw-bytes or Schema item in a make() list crashes construction #1312 builds: valid raw items build in all 6 protocols probed. bytes(cls(**kw)) equals the maker's make().pack() across about 6,000 fuzz successes, with 0 differences.
• fix(protocols): resolve a str enum name against its own make() argument rather than raising on a shared name #1308 fallback: the per-argument resolution is lru_cached. It falls back correctly for exec'd classes, .pyc-only modules, -OO, a dedent SyntaxError, aliases and functools.partial.
• Tests: the legs passed at 503/0 and 907/0, and tests/project gave 389 passed. Fixture info matches main. The revert proof gives 16/2.

[JarryShaw]

Round-2 verdict on f345fe84e: GOOD TO GO. I re-ran my round-1 probe on this head myself. The head is rebased onto 4849d16e4.

• Blocker fixed. IPv4(options=[b'\x02', b'\x94\x04\x00\x00'], protocol=253) now raises ProtocolError: IPv4: malformed raw item given to make(), chained from the parse error. It no longer raises AttributeError. The except is narrowed to BaseError and struct.error. A seeded fuzz of 1,500 inputs across all 8 protocols gives 0 non-library exceptions, where the old code gave 323 AttributeErrors.
• Valid raw items still build. IPv4 Router Alert gives 24 B, a TCP MSS gives 24 B, and a bare MH() gives 8 B.
• Hot path. The schema walk now runs only when make() was given bytes or a Schema. The cost over main falls from about 11% to about 4–6%: 89.4 µs to 93.0 µs for TCP(), and 109.9 µs to 116.0 µs for IPv4(protocol=253), both measured at load ≈6.
• Loop-variable regression fixed. It now falls back to the all-arguments check, so x = b; self._make_index(x) behaves as it does on main. The 12 hip.py sites are named in the body.
• Tests and checks. The previous head fails 8 of the new tests. The legs give 529/0 and 918/0. Fixture info is identical to main.

One residual gap, stated in the PR body. A tuple-form option whose maker returns raw bytes would not set the flag. No such maker is known in-tree.

[github-actions]

Coverage: 89.31% (unit tier, Python 3.14, f345fe84e, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	4	93.55%
pcapkit/const	18777	966	2348	869	90.90%
pcapkit/corekit	1949	83	606	19	94.99%
pcapkit/dumpkit	189	1	68	1	99.22%
pcapkit/foundation	2443	102	852	39	94.39%
pcapkit/interface	112	7	40	5	92.11%
pcapkit/protocols	16076	187	4216	158	98.26%
pcapkit/toolkit	539	74	168	3	84.58%
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.