perf(cli): count a cluster's members once, not once per member (#175)

[xping-admin]

Closes #175.

SharedFailure counted each cluster member's failures with a linear scan over the whole cluster, inside the loop over the cluster's members — O(|Fingerprints| × |Failures|) with an ordinal string comparison in the inner loop, once per cluster.

The scaling factor is the payload the kind exists to report. A broken fixture takes out every test in an assembly, so the members and the failures grow together: the pathological input is precisely the successful case. The default window reads 20 sessions, which bounds it. --since reads up to 1000.

The counts move into the index

SignatureGroup gains FailuresByFingerprint, counted in the same walk of the same list that already produces Fingerprints, MaxTestsInOneSession and SessionCount — so the counts are free there, and there is no longer a second place that could decide what "this member's failures" means. SharedFailure reads them.

The loop keeps its shape deliberately. references and members are appended together only when the fingerprint resolves, and FindingSubject.Group is built from the first while the evidence is built from the second; hoisting the whole loop to pre-group would break that pairing.

Spread was the other half

Same list, also once per cluster. It ordered every failure under a four-key comparator — materialising ExecutionId.ToString("N") for each one — to read three entries off the front.

It now keeps the earliest failure per key in one pass and orders those, which is the same answer: the entries of the ordered list that introduce a new key are exactly the per-key minima, and no two minima can tie (a tie needs the same run, attempt, test and execution id, which is one execution and so one key). The top-up branch, reached only below three distinct keys, selects into a bounded shortlist instead of sorting.

The last comparison key stays the hex form rather than Guid.CompareTo, which orders differently. It is now formatted only for two attempts of one test in one run.

Measured

500 runs × 1000 tests, all failing on one signature:

	index build	cluster analysis
before	1842 ms	22 960 ms
after	1843 ms	75 ms

Emitted evidence byte-identical, at that size and at 200 × 500. At the issue's headline 1000 × 2000 the analysis is 339 ms.

Worth a reviewer's attention: what now bounds that store is SignatureIndex.Build signing two million failed executions — a regex pipeline plus SHA-256 each — at 6.7 s. That is linear rather than a blow-up, so it is a separate question from this one, but it means the issue's "report in seconds at 1000 × 2000" is not reached by this change alone.

Two corrections to the issue

• context.Tests.ExecutionsOf is a dictionary probe (TestIndex.cs:85), so the Count(...) was the only quadratic term in that loop. The issue left it open.
• There are no golden files in this repo. The acceptance bullet about byte-identical golden output had no artefact behind it; ReportEnvelopeTests compares two runs against each other, which catches nondeterminism but not a changed value. The byte-comparison above was run against the pre-change provider instead.

The issue's two "noted only so they are not fixed unnecessarily" items — RetryProvider.MaskedExemplars and DurationProvider.SessionMedians — were both re-checked and both left alone.

Verification

dotnet test tests/Xping.Cli.Tests — 910 pass, 2 new.

Nothing in the suite asserted ClusterMember.Failures at all, and nothing pinned which exemplars a cluster picks rather than how many. Both new tests pass against the pre-change provider, which is what makes them evidence that the selection is unchanged rather than merely self-consistent.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TMQcZLBGFTd8wz3mcfn9vc

[Copilot]

🟡 Changes recommended

The repeated-key exemplar top-up path lacks an exact output-order regression test.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Optimizes shared-failure analysis while preserving evidence ordering.

Changes:

• Precomputes failure counts per fingerprint.
• Replaces full exemplar sorting with bounded selection.
• Adds count and ordering regression tests.

File summaries

File	Description
SignatureIndex.cs	Indexes per-fingerprint failure counts.
FailureModeProvider.cs	Uses indexed counts and optimized exemplar selection.
FailureModeProviderTests.cs	Tests member counts and distinct-key ordering.

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/Xping.Cli/Report/Indexes/SignatureIndex.cs	97.65% <100.00%> (+0.07%)	⬆️
.../Xping.Cli/Report/Providers/FailureModeProvider.cs	96.16% <100.00%> (+0.41%)	⬆️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.