refactor(corekit,protocols): normalise the four sentinel objects to SCREAMING_SNAKE (#937)

[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 — 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

Implements #937: NoValue -> NO_VALUE, _Absent -> ABSENT (type _AbsentType -> AbsentType), per the owner's ruling to accept the breaking change with no backport. ABSENT/AbsentType stay out of every __all__; privacy is now documentation-only (sentinels.rst marks them private explicitly, per the ruling). conventions.rst gains the object-naming rule next to the existing type-naming rule. test_sentinel_exports_unit.py's _expected_type_name collapses to one mechanical rule, and a new test pins that every object is SCREAMING_SNAKE (confirmed failing on 382375811).

Also fixes the 3 NoValue references in tests/protocols/internet/test_mh_unit.py (:1869, :1875, :1879) — nothing else in that file. It was initially treated as off-limits alongside its zero-reference source, mh.py (owned by #935), but the test itself was never actually measured; CI's 5 Python-version legs caught it, failing identically on test_mh_padding_option_schema_sizes_itself.

Two build-tooling notes worth keeping on record:

• isort reordered 5 files' imports from Field, NoValue to NO_VALUE, Field: order_by_type sorts CONSTANT-cased names ahead of Class-cased ones, so a pure casing change alone can reorder imports.
• pylint's cyclic-import (R0401) output is nondeterministic run-to-run: two runs of the unmodified 382375811 tree produced 236 differing lines, so R0401 churn can never be pinned on a diff. Every other diagnostic (5,815 lines) was byte-identical against baseline.

[JarryShaw]

NEEDS CHANGES at 8c5feda9c — six red marks, one root cause, and the partition error is mine, not this PR's.

tests/protocols/internet/test_mh_unit.py:1875: ImportError:
  cannot import name 'NoValue' from 'pcapkit.corekit.fields.field'

All five Python 3.10–3.14 legs fail identically on MHUnitTests::test_mh_padding_option_schema_sizes_itself, and Required checks passed fails downstream of them. 54 other checks pass — Lint, Compat, the whole Engines matrix, Integration, parity, Analyze, Changelog drift, safety.

Why it happened. I declared tests/protocols/internet/test_mh_unit.py off-limits to this PR's worker because PR #940 owns it, and I justified that by measuring "zero sentinel references in mh.py" — the source file, not the test file. The test file has three: :1869 in a docstring, :1875 the import, :1879 the use. So the partition was verified against the wrong file, and the worker did the right thing by stopping and reporting rather than editing across the line I drew.

Resolution: this PR takes those three lines. A rename cannot land while leaving an import of the old name behind, and the three references are purely mechanical — NoValue → NO_VALUE, including the :data: role in the docstring.

Consequence for #940, which also edits this file (+94 lines): the two will conflict on merge. Whichever lands second rebases — normal, and cheaper than splitting the rename.

Verified good otherwise, and two findings worth recording:

• Reference counts re-derived and matching: NoValue 109/21 files, NoValueType 55/13, _Absent 41/5, _AbsentType 26/5.
• ABSENT/AbsentType are documented as expressly private on sentinels.rst with the ruling quoted, rather than omitted — the right call, since omission would just be underscore-hiding by another name. Both confirmed out of pcapkit.protocols.protocol.__all__ and pcapkit.corekit.sentinels.__all__.
• isort initially broke five files: NO_VALUE sorts before Field/FieldBase under order_by_type, which the old CamelCase name did not. Fixed, now clean — worth knowing that a case change alone can reorder imports.
• pylint's cyclic-import output is nondeterministic: a double run on an unmodified baseline produced 236 differing lines with zero code change. So R0401 churn is never attributable to a diff. Non-cyclic diagnostics are byte-identical here (5815 lines each), mypy byte-identical at 321/38.

[JarryShaw]

GOOD TO GO at 7090f84bd — Fable cross-review (author was Sonnet), with its one blocking finding already resolved by the head it did not see.

The review measured 8c5feda9c and returned NEEDS CHANGES on exactly one defect: the three live NoValue references at tests/protocols/internet/test_mh_unit.py:1869, :1875, :1879, demonstrated as ImportError: cannot import name 'NoValue'. That was my partition error — I declared the file off-limits on the strength of measuring the source mh.py, never the test — and it is fixed at 7090f84bd, whose delta is exactly those three lines and nothing else.

Independently confirmed by the review, each with its own measurement:

• Completeness: tree-wide grep -rnw over pcapkit tests docs examples util. Every other old-name hit is deliberate historical prose or the unchanged type NoValueType. Zero NO_VALUEType botches — that substring trap was the one I most expected to bite — and the _ABSENT hits are verbatim ruling quotes.
• Privacy: sentinels.__all__ == ['NULL','NO_VALUE','NO_DEFAULT'], field.__all__ == ['NO_VALUE','Field'], protocol.__all__ == ['ProtocolBase']; the test at test_sentinel_exports_unit.py:289 pins object, type and import *.
• Runtime: NULL/NO_VALUE/ABSENT falsy, NO_DEFAULT truthy, repr(ABSENT) == '<absent>', no NoValueType.__repr__, BytesField(length=4).default is NO_VALUE with del restoring it, and no alias — the old names raise ImportError, per the ruling.
• Anchors intact and in order at 25 / 142 / 308 / 776; _expected_type_name collapsed to the mechanical rule with strength kept by the new regex test; mypy 321/38 and pylint 8.67, exit 30, both at baseline; isort clean.

It also corrected my brief, and I verified the correction. I told it all four sentinels have re-export shims in pcapkit.protocols.protocol; that module has only ever shimmed the Absent pair — from pcapkit.corekit.sentinels import _Absent, _AbsentType on main, ABSENT, AbsentType on the PR tree. My error, not the PR's.

UNVERIFIED and worth stating: the coverage delta was not measured (budget), and test_every_sentinel_object_is_screaming_snake could not be shown to fail individually on 382375811 because the whole file errors at collection there — which subsumes it but is weaker evidence than a targeted failure.

Unpublished and awaiting you. One coordination note: #940 also edits test_mh_unit.py, so the two will conflict at merge and whichever lands second needs a rebase — a consequence of my partition error, not of either PR.