[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
• Changelog entry: N/A — centralised in chore(docs): centralise changelog entries for unreleased versions #2483

What is the purpose of your pull request?

• refactor — changes neither behaviour nor performance

---

Description of your pull request and other information

The rest of #2676, after the registry-code bullet, which #2594's rework takes.

• Fixed-size octets: a caller's octets for a fixed-size field go through _make_octets and raise ProtocolError at any other width, instead of being padded or cut. This covers the checksums of IPv4, MH, HIP, ICMP, IGMP, IPX and OSPF, the CALIPSO checksum, MH cookies and keygen tokens, HIP opaque and payload data, PCAP-NG ZigBee keys, IPX addresses, and the reserved and unused octets of the IPv6 extensions, MH, HIP, ICMP, IGMPv3, OSPF and IP_Pseudo. A source sweep in the tests finds every such field a maker writes, so a new one fails without a row.
• IPv4 timestamp option: a slot count past what the option holds raises, instead of being clamped. The default is derived from the form, so both defaults build the same octets as before.
• Parsed models rebuild as captured: a parsed MH message, IPv6-Route type, IPv6-Frag header or OSPF authentication model keeps its octets as captured. Only a caller's value is checked, so a capture cut inside them still rebuilds.
• Docs: protocol.rst documents the shared value helpers.
• MH ipv4_dhcp.mode=True is unchanged: it is a one-bit field, which takes a bool.

Breaking: released makers that padded or cut a fixed-size value now raise.

Closes #2676

[JarryShaw]

NEEDS CHANGES at 9ee57bdf8, from a Sonnet cross-review. Caller values, defaults, the timestamp counts and the parse output are all confirmed, and captures are identical to main. What has to change:

1. The rebuild exemption lets a caller-built model through. It treats any data model as a rebuild: MH msg_reserved, IPv6-Route type 0 and type 2, and OSPF crypto auth. So a hand-built model with a wrong-width reserved field is padded or cut instead of refused. Follow IPv6_Frag.from_data: give the maker a full-width value, let replay_short_read trim, and check every value. At the very least, key the exemption on __short_read__.
2. No test pins the exemptions. Removing any of them fails nothing. Add direct-parse cut tests for each exempted site, and a test that a wrong-width model is refused.
3. CI: gate ast.unparse (test(project): the #1878 engine-stub tests call ast.unparse, which Python 3.8 lacks #1899), and add the new test to project-tests in unit-tests.yml.

[JarryShaw]

Pushed the fix round 9ee57bdf8→6ca69b2ec, rebased onto 57a6c366d.

• Every value is now checked: _make_octets zero-fills a short value only when the model carries the built-in __short_read__ record. A caller-built model is checked like any other value.
• Each exempted site is pinned: direct-parse cut tests cover MH, IPv6-Route 0 and 2, IPv6-Frag and OSPF, and wrong-width models are refused.
• CI: ast.unparse is replaced, and the docs check moved under tests/project.

A Sonnet re-review is running.

[JarryShaw]

NEEDS CHANGES at 6ca69b2ec, from a Sonnet re-review. Most of it is confirmed:

• every cut capture at every offset rebuilds exactly;
• the CI fixes and rebase resolutions are right;
• captures are identical;
• 284 modules pass.

One hole remains: the zero-fill fires whenever the model carries any __short_read__, not only one in the reserved field. So editing msg_reserved (MH) or reserved (IPv6-Route 0) to a short value, on a model cut somewhere else, is padded instead of refused.

1. Tie the fill to the record naming this field, or to a cut that falls at or inside it.
2. Add Frag to the wrong-width test. An unconditional Frag fill survives today.
3. Add the "record elsewhere, reserved edited short" case for MH and Route.

[JarryShaw]

Pushed the fix round 6ca69b2ec→ea9935d45, still one commit on 57a6c366d.

• The fill is now Protocol._fill_short_read. It zero-fills a short value only when this model's own __short_read__ record shows the cut ended inside the field or before it. A value cut elsewhere, or of any other width, still goes to _make_octets and is refused. Both of the re-review's repros now raise ProtocolError.
• This covers the 8 MH msg_reserved messages, IPv6-Route types 0 and 2, IPv6_Frag and OSPF. _make_msg_brr needs no fill, because its reserved field is an integer.
• New cases cut past each field and shorten it. Run against 6ca69b2ec, they fail 12 subtests.

A Sonnet re-review is running; review: running.

[JarryShaw]

NEEDS CHANGES at ea9935d45, from a Sonnet re-review that I reproduced. The fix for the previous finding holds: _fill_short_read fills only on this model's own record, and both earlier repros are refused.

• New regression: an OSPF cut at 16 or 17, inside the cryptographic-authentication reserved octets, rebuilds from info but raises ProtocolError ("must be 2 octets") from info.to_dict(). OSPF._make_data fills only when auth is a Data_CryptographicAuthentication, and to_dict() turns it into a mapping. Both cuts rebuild on main.
• Fix: apply the fill when auth is a mapping too, and extend test_ospf_cut_inside_authentication_reserved_octets to the to_dict form.
• The full regression legs ran out of time in the review, so they still need a complete run.

[JarryShaw]

Pushed ea9935d45→d57c0a45f, rebased onto 596dcfaf8, still one commit.

• The OSPF fix: _make_data fills any mapping auth, so cuts at 16 and 17 rebuild from info.to_dict() as well as from info. The test now covers the to_dict form and fails without the fix. Sweeping every cut at all 12 fill sites, through both forms, is clean.
• The rebases: these resolved docstring conflicts only. refactor(protocols)!: make() takes an enum member or an int, with no keyword groups (#2594) #2692's maker shapes, refactor(corekit)!: move the schema helpers protocols call into corekit.schema (#2674) #2704's paths, fix(internet)!: keep MH option 6 prefixes with host bits and option 27 times past 9999 (#2702) #2709's MNP and HNP text and fix(protocols): refuse or build what the TCP, HIP and MH makers mishandled (#2708) #2724's HIP changes are kept, each with this PR's width clause added. The range-diff shows no change in behaviour.
• Tests: 317 modules pass, covering internet, the protocols root and project.

A Sonnet re-review is running; review: running.

[JarryShaw]

GOOD TO GO at d57c0a45f, from a Sonnet re-review.

• OSPF: every cut from 1 to 23 rebuilds exactly through both info and to_dict. Reverting only the mapping fill fails the test at 16 and 17.
• Other fill sites: every one is clean through both forms, and both earlier repros are still refused. _fill_short_read is byte-identical to the version reviewed last round.
• The rebases: none undoes merged work. Every removed line is a docstring or a pass-through, replaced by a width check. The changes from refactor(protocols)!: make() takes an enum member or an int, with no keyword groups (#2594) #2692, refactor(corekit)!: move the schema helpers protocols call into corekit.schema (#2674) #2704, fix(internet)!: keep MH option 6 prefixes with host bits and option 27 times past 9999 (#2702) #2709 and fix(protocols): refuse or build what the TCP, HIP and MH makers mishandled (#2708) #2724 are intact.
• Tests: 419 modules pass.