[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 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?

• ci — workflows or build tooling

---

Description of your pull request and other information

Phase 1 of #1958, implementing rulings A and B.

• tests/_tiers.py gives every tests/**/test_*.py module one tier, ALL or LATEST.
  • A module-level __tier__, read by AST, wins. A bad value, a second assignment, or one that is not at top level fails.
  • Otherwise the module is LATEST if its run does not vary by interpreter. This is ci(unit-tests): run platform-independent project pins once per push (#1954) #1956's rule, generalised to all of tests/.
• Real sweeps are marked LATEST; hand-picked round trips are marked ALL. Every marker carries a #: Test tier (#1958): <reason> comment.
• The *roundtrip* globs are gone.
  • Legs off latest (supported, KTLO, experimental) ignore latest_tier_paths().
  • integration off latest and both pypcap-parity legs run fixture_tier_paths('ALL'), so parity no longer runs the sweeps against ci(python): the GraalPy experimental leg reaches 26% of the suite before its 25-minute cap #1732.
• Unchanged: the latest leg runs both tiers under coverage, the pins stay in project-tests, and gate and unittest-ordering keep their selections.
• The tier guard, the ci(python): the GraalPy experimental leg reaches 26% of the suite before its 25-minute cap #1732 and ci: stop repeating text-only PR pin tests on every main unit leg #1954 pins and _dependency_gates.py check:
  • the partition;
  • the marker;
  • the workflows' agreement.
• workflows.rst gains a "Test Tiers" section.

Every collected test id lands in exactly one tier, with none lost or duplicated. The tier lists are identical under 3.8, 3.10, 3.12 and 3.14.

Refs #1958

[JarryShaw]

GOOD TO GO at d6fc7d507. A Sonnet review recollected the whole tree: 5511 ids, every one selected by exactly one job's tier. Every ALL id runs off latest and on latest. No module testing interpreter-specific behaviour (compat, enum, io, recursion limits) landed in LATEST. latest_tier_paths() is byte-identical on a real 3.8 and on 3.14, and every changed file parses as 3.8. A failed tier list fails the leg loudly; experimental warns. Doctoring the workflows or a marker trips the guards. It merges cleanly with main (including #1964's new module, which resolves to ALL).

One question for you, on ruling B. The four *_edge_roundtrip_unit modules and test_short_read_roundtrip_unit stay LATEST as tables of every #1202 edge case. Their rows are hand-written, though, so they could count as hand-picked and move to ALL. Which do you prefer?

[github-actions]

Coverage: 90.07% (unit tier, Python 3.15, d6fc7d507, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	2	20	2	96.77%
pcapkit/const	19597	881	2618	867	92.11%
pcapkit/corekit	2504	70	792	36	96.54%
pcapkit/dumpkit	276	3	100	3	98.40%
pcapkit/foundation	2916	164	1016	34	93.59%
pcapkit/interface	138	7	52	5	93.68%
pcapkit/protocols	17411	224	4580	189	98.10%
pcapkit/toolkit	872	177	298	8	77.18%
pcapkit/utilities	522	2	158	4	99.12%
pcapkit/vendor	4431	2341	1018	156	43.64%

Per-file detail: the coverage-html artifact of this run.

[JarryShaw]

I merged without giving decision. What's your recommendation?

[JarryShaw]

My recommendation is to split them:

• Move the four *_edge_roundtrip_unit tables to ALL (application, internet, link, transport). Their rows are hand-written edge cases that exercise parsing and building, which is ruling B's purpose: quality assurance on every interpreter. Together they take about 15 s on 3.15.
• Keep test_short_read_roundtrip_unit on LATEST. It cuts each case at every offset, so it is a real sweep whose cost grows with its inputs (8.3 s).
• Keep test_edge_roundtrip_deadline_1736_unit on LATEST. It tests the harness's deadline, not pcapkit.

If you agree, it is a four-marker follow-up.

[JarryShaw]

agreed.