Attribute skipped tests per file, and re-lower the dead-key ceiling (#132)

[juemerson-at-purestorage]

Two test-only gate fixes. Both are cases where a gate had gone slack and was reporting protection it no longer provided.

1. Dead-key ceiling: 85 → 83

PR #134 fixed #119, which removed the two source.name records for Invoke-PfbNetworkPing and Invoke-PfbNetworkTrace — one of them a severity: WRONG-RESULTS entry. The committed report went 85 → 83.

Nothing red, because 85 is a ceiling and a drop is the direction it calls better. That is exactly why it needs re-lowering by hand: left at 85, the gate silently tolerates two brand-new dead keys.

It stays a ceiling, not a pin. Dead keys legitimately fall as fixes land, and a pin would red every such fix — making the gate a tax on doing the right thing.

2. Per-file skip attribution, replacing the global ceiling (#132)

Tests/coverage-baseline.psd1's per-edition MaxSkipped becomes ExpectedSkips: an exact expected skipped-test count per test file.

The ceiling carried deliberate headroom, and the headroom was the defect rather than the mitigation. It did not prevent false reds; it converted an attributable red on the PR that moved the number into an unattributable red on a later, innocent PR — and it hid a real coverage regression of up to ceiling - measured tests, which is the #63 failure shape merely bounded in size.

The arithmetic behind that is exact rather than hypothetical: a 5.1 run measured 292 against a ceiling of 268, and six of the +40 belonged to #120 two days earlier, which had slack and passed without touching the baseline. The next branch across the line was billed for all forty. The five raise notes in the winps51 block are retained verbatim as the audit trail — by the last one they are visibly diagnosing the mechanism rather than the branches ("268 was already stale against main before either branch existed").

The implementation detail that decides this works

Attribution reads $Result.Containers, not $Result.Tests, and the difference is load-bearing: a file whose #requires stops it before discovery contributes no entries to .Tests at all, but still appears in .Containers with every count at zero. So .Containers is the only place that failure is visible. Probed directly on both editions (Pester 6.0.1 under pwsh 7, 6.0.0 under Windows PowerShell 5.1) rather than assumed.

Rails that come with the map

• No test file may run empty — contributing neither an executed nor a skipped test. This is the CI silently skips ~23% of the test suite, including the absolute-path regression guards #63 shape below the granularity RequiredDescribes reaches, which only covers hand-listed blocks.
• No undeclared file may skip.
• A declared file that stops running at all is a violation, not a stale entry to delete. The tempting reading is the one that turns a regression into a tidy-up.
• Per-file skips must reconcile with the run's own total, so a container the walk misses is a red rather than a quietly smaller number.

Where the numbers came from

Seeded from run 32392093324 on main@98c7a16, whose Detailed output attributes every skip to a file — so no local aggregate-suite run and no deliberately-red first push was needed to measure it:

Leg	Skips	Files
pwsh 7 (ubuntu, windows, macos — all three agreed exactly)	2	2
Windows PowerShell 5.1	297	19

Both sums reconcile with that run's own totals, which is the control proving the parse. The three pwsh legs agreeing is what makes one shared pwsh7 map correct rather than a Windows-only measurement generalised. Only files that actually skip appear, so the map is 21 lines rather than a 199-file census.

Verification

• Scoped Pester, both editions: 32 passed / 0 failed / 0 skipped, Container ok, on the three touched test files.
• Tests/CiCoverageGate.Tests.ps1 grows from 7 to 15 gate tests, including the direction a ceiling structurally could not see: a declared file skipping fewer tests than expected. Also asserts MaxSkipped is absent from both blocks, since a leftover key nothing reads is a no-op masquerading as a gate.
• End-to-end against the real baseline: it passes when fed a stand-in built from the exact per-file numbers above, and reds on each of three mutations — one declared file moving by +1, an undeclared file starting to skip, and a file running empty.

Live-FlashBlade verification does not apply. The diff is Tests/ and scripts/ only, leaving Public/, Private/, PureStorageFlashBladePowerShell.psd1 and PureStorageFlashBladePowerShell.psm1 entirely untouched, so it cannot change what goes on the wire.

No version bump and no CHANGELOG entry, per the usual convention.

🤖 Generated with Claude Code