[JarryShaw]

• Searched for similar pull requests
• Followed the coding style
• 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?

• test — tests only

---

Description

Closes #2160.

• Cause (measured, not reproduced): the old bound never failed locally, in 30 fresh PyPy 3.10 runs, under 5M–40M-object heaps, or in a long tests/protocols process. PyPy's GC hooks show each deep parse triggers 2–3 full-collection steps where a short parse triggers none. That cost grows with heap size: deep IPv6 took 8.1, 13.6 and 25.0 ms at 0, 10M and 40M live objects, and 9.4–10.0 ms at every size with the collector off. A CI worker's heap is the likely source of the 125 ms in run 40161694146. Runner CPU steal is the other candidate.
• Fix, test only: best() runs gc.collect() and disables the collector while timing, then restores it, as timeit does. The bound deep < 4 * short + 0.05 is unchanged.
• Results with the fix: PyPy's worst ratio was 1.74 fresh and 1.23 with a 40M heap. The timing test passed 30/30 on PyPy and 30/30 on CPython.
• Regression guard: new fake-clock TimingBoundTests check that the collector's cost stays out of the timing, that the GC state is restored even when the parse raises, and that a linear or quadratic per-layer cost still breaks the bound. Without the fix the first test fails (0.1375 != 0.0125).
• Tests: the whole file passes on CPython 3.14 and PyPy 3.10 (16 tests, 53 subtests each). tests/test_tier_guard.py: 133 passed.

Not in scope: the bound is loose. A quadratic walk past the limit is caught for IPv6 but not for IPv4 or 802.1Q, where the 0.05 s slack dominates. Tightening it needs measurements across the CI matrix, so it is tracked separately.

[JarryShaw]

GOOD TO GO at 185daec00 (Sonnet review).

• GC state restored: best() captures gc.isenabled() before collecting and re-enables in finally only if it was enabled. This holds on a normal return, on an exception from parse, and when the GC was already off on entry. test_the_collector_is_left_as_it_was covers both entry states. A mutant that always re-enables fails it, and the addCleanup stops any leak into later tests.
• The guard fails without the fix: removing gc.disable() gives 0.1375 != 0.0125. The guard drives the real module-level best() and bounded(), with only the clock injected.
• CPython is unchanged: bounded() is deep < 4 * short + 0.05, the same as the old inline check, and the parse is still inside the timed span.
• Tests: CPython 3.14 ran the file 3 times, 16 passed each time; tests/test_tier_guard.py 133 passed. The file is tier ALL.
• UNVERIFIED: the PyPy 3.10 measurements, because no PyPy was available to the reviewer. The PR body says the cause was measured but not reproduced.

[github-actions]

Coverage: 91.55% (unit tier, Python 3.15, 185daec00, Unit Tests run success)

Package	Statements	Missed	Branches	Partial	Cover
pcapkit (top level)	104	4	20	2	95.16%
pcapkit/const	19599	880	2618	866	92.12%
pcapkit/corekit	2589	75	816	37	96.48%
pcapkit/dumpkit	282	3	102	3	98.44%
pcapkit/foundation	2971	23	1036	25	98.80%
pcapkit/interface	138	4	52	3	96.32%
pcapkit/protocols	17418	219	4574	185	98.14%
pcapkit/toolkit	869	36	298	10	95.20%
pcapkit/utilities	504	7	146	5	98.15%
pcapkit/vendor	4432	1988	1018	167	51.87%

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

[JarryShaw]

Ready to merge at 185daec00. Merging closes #2160.

• Review: Sonnet GOOD TO GO. The GC state is restored on every path, and the guard fails without the fix.
• CI: all 62 CheckRuns passed, none failed or pending. Required checks passed is SUCCESS, and the KTLO legs show no warnings, including the pypy3.10 leg that raised test: the frame-layer-limit timing bound flakes on PyPy 3.10 (KTLO) #2160.
• Ordering: I merged this PR into main and ran tests/protocols plus the root modules under the plain unittest ordering leg: 1903 tests, 0 failures, 0 errors.
• Conflicts: none. Only the one test file changed.