ci(unit-tests): install DPKT so the dpkt-gated tests actually run

[JarryShaw]

Please follow the guide below

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort) — N/A, YAML-only change, no Python touched
• make test passes, and a test case covers the change — forbidden here (full suite); see test evidence below instead
• 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
• feat
• perf
• refactor
• test
• ci
• chore

Description

Fixes #729. dpkt is declared only in the DPKT/all extras; no unit-tests.yml install line named it (.[test] for test, .[test,Scapy] for integration/gate), so all HAS_RUNTIME and HAS_DPKT-gated tests skipped on every job/version, forever, reading as passes.

AST-measured (still accurate after #725/#731 shifted the issue's line numbers): 28 methods, 6 files — 20 test_dpkt_unit.py, 4 test_misc.py, 1 each test_extraction.py/test_core.py/test_engine_parity.py/test_engine_runtime.py.

Added DPKT to all three install lines, on every existing leg (not one representative leg): dpkt 1.9.8 installs and imports cleanly on 3.10–3.14 (measured, throwaway venvs), it adds no new job/leg — unlike Scapy, which already sets the "extra on every leg" precedent on these same lines — and singling out one leg would be an inconsistency, not a wall-clock saving (#715 is about job count, not per-job extras). HAS_RUNTIME's four packages are core project.dependencies, always installed already, so DPKT alone is sufficient — no separate fix needed there.

test_reassembly_engine_parity.py also imports HAS_DPKT but only branches internal assertions on it; it isn't skipUnless-gated and isn't one of the 28.

Test evidence (throwaway /tmp venvs, repo .venv never touched; pcapkit.__file__ confirmed pointing at this worktree): before (.[test] / .[test,Scapy], no dpkt): 26 passed/28 skipped + 6 passed/3 skipped. After (+DPKT): 62 passed, 1 skipped — the 1 remaining skip is an unrelated pyshark not installed. All 28 previously-skipped dpkt methods now pass.

[JarryShaw]

✅ GOOD TO MERGE @ 28582b25c — re-derived 28/6 independently, measured all 28 now execute, dpkt wheel-installs on all five Pythons, 15 required contexts intact.

[JarryShaw]

✅ GOOD TO MERGE @ 28582b25c — re-derived 28/6 independently, measured all 28 now execute, dpkt wheel-installs on all five Pythons, 15 required contexts intact.

Independent cross-review on a different model from the author. Every figure below is one I obtained myself, not one taken from the PR body.

Claim	Verdict	Evidence I obtained
28 gated methods across 6 files	✅ exact	Own AST pass resolving class-level and method-level skipUnless gates: 20 / 4 / 1 / 1 / 1 / 1 — identical split. The two other HAS_DPKT files (tests/integration/_helpers.py, test_reassembly_engine_parity.py) carry no gate, as the body says
All 28 now execute, dpkt is the sole blocker	✅ 28/28	Those exact 28 node-ids, fixtures regenerated: dpkt hidden → 28 skipped; dpkt visible → 28 passed … 39 subtests passed. Only variable changed was find_spec('dpkt')
Every gated file is reached by a leg that now has DPKT	✅ no gap	test ignores tests/integration + *_runtime.py, so it carries 24; fixture_tier_paths() hands integration the other 4 — tests/integration plus two test_dpkt_unit.py node-ids. 24 + 4 = 28. Both changed lines are load-bearing
CI corroborates, zero collateral	✅	Python 3.12 1542→1566 passed / 116→92 skipped (+24 / −24); Integration 3.14 143→147 / 31→27 (+4 / −4). Exactly the partition above and nothing else moved
dpkt installs on 3.10–3.14	✅ measured, all 5	dpkt-1.9.8-py3-none-any.whl (194 kB, pure Python, no sdist build) on all five test legs and all five Integration legs
15 required contexts survive	✅	name: lines byte-identical to main (line numbers only shifted); all 15 green, rollup SUCCESS, mergeStateStatus: CLEAN, rebased on 55513f69e as one commit
HAS_RUNTIME needs no fix	✅	RUNTIME_DEPS = ('tbtrim','aenum','chardet','dictdumper') — all four in project.dependencies
Template	✅	Four-item checklist, exactly one type (ci), Fixes #729.; both standing exceptions worded correctly
Coverage rule	n/a	Diff is one file, .github/workflows/unit-tests.yml; no test or library file touched, so the rule does not engage. pytest-cov still absent; I installed nothing anywhere

On the matrix legs I agree with all five, not one. #715 was job count (~130 min/push of duplicated legs); this adds no job, and a 194 kB pure-Python wheel costs about a second per leg. Measured wall-clock stays inside pre-existing variance (test legs 411–578 s against a 559 s baseline and a 45-min timeout). One representative leg would leave 28 tests on single-version coverage while 1566 get five — and pcapkit's dpkt adapter is byte/struct-level work, which is precisely what a version matrix is for. Scapy already sets the every-leg precedent on these same lines.

Caveats and disagreements, none blocking:

• The gate job's changed line is unexercised by this PR: if: inputs.gate-only == true, so it reported skipped with steps: []. Correct by inspection, but verified by nothing here.
• The body's headline evidence conflates two variables. "before … 26 passed/28 skipped + 6 passed/3 skipped; after 62 passed, 1 skipped" is a delta of 30 resolved skips, but only 28 are dpkt-gated — the other 2 are scapy-gated tests that resolved because the "after" venv also gained Scapy. Isolating dpkt alone across the same six files gives 35 passed, 28 skipped → 63 passed, 0 skipped, a delta of exactly 28. The conclusion is right; the evidence as written does not isolate it.
• The new integration comment says "the two dpkt-gated methods living under tests/integration/". That leg actually carries four — the other two being test_dpkt_unit.py::DPKTEngineParityTests::test_tcp_reassembly_matches_the_default_engine and ::DPKTTCPHeaderSplitTests::test_every_tcp_frame_of_the_sample_capture_holds_the_invariants, which fixture_tier_paths() routes there because they need test.pcap. Comment-only, and the code is right.
• The body drops the docs commit-type box (template lists eight, body seven).
• CI runs pytest -q with no -r, so skip reasons never print: "0 occurrences of dpkt not installed" would read the same whether or not the fix worked. The skip delta is the only usable signal there.
• Why this survived so long: Pipfile [dev-packages] already had dpkt = "*", so local make test always ran these 28 and only CI was blind. The fix is at the right layer.

✅ GOOD TO MERGE @ 28582b25c — re-derived 28/6 independently, measured all 28 now execute, dpkt wheel-installs on all five Pythons, 15 required contexts intact.