test(anomaly): guard fixture search work both ways; make -bench . finish

[dborup]

Summary

Benchmark hygiene in internal/anomaly, following the independent review of PR #108. Only *_test.go files change, and there are no production code changes. The library is still not wired into any binary.

1. The work guard now covers the fixture rule and checks both ways. TestPeriodicSearchWorkIsBounded did not see the +41 % rise in visited gaps that fix(anomaly): periodic refinement, mixed-burst labels and censored-episode confidence #108 caused on the fixture rule's alternating shape (72.96 → 103.0 gaps per pulse). It also could not see a drop in work, which would mean lost search coverage.
2. BenchmarkBytesPerStream never finished. It repeated a whole heap measurement b.N times with the timer stopped, so go test -run '^$' -bench . ./... did not complete. This predates fix(anomaly): periodic refinement, mixed-burst labels and censored-episode confidence #108.

1. Work guard (periodic_bench_test.go)

Before:

• The guard ran only maxPeriodicRule() and worstPeriodicRule(). The rule behind BenchmarkPeriodicSearch's fixture/... cases (experimentalPeriodic()) was not covered.
• measuredSteps was keyed by shape/MinCoverage. The fixture and max rules both have MinCoverage 0.6, so that key could not tell them apart.
• The check was one-sided: it failed only above +25 %.

Now:

• Rules: the guard runs fixture, max and worst. measuredSteps is keyed by rule/shape and holds master's actual totals (18a13264, 4 × 256 pulses per rule and shape).
• Two-sided margin: it fails above +25 % (more work), and also below −25 %. The error for the lower bound says the search visits markedly fewer gaps, which means lost search coverage, and that a deliberate optimisation should update measuredSteps.
• Comment: it now states that the counters count only visited gaps. Work per gap, such as the periodRange memo, is not visible to them. TestRangeMemoIsNeverStale covers the memo, and only BenchmarkPeriodicSearch measures the CPU.

rule/shape	previous value (key)	new measuredSteps (master 18a13264)
fixture/periodic	not guarded	33,289
fixture/alternating	not guarded	104,271
fixture/jitter	not guarded	54,979
fixture/missing	not guarded	34,850
fixture/bursty	not guarded	82,025
fixture/noise	not guarded	67,588
max/periodic	236,631 (periodic/0.6)	236,631
max/alternating	1,466,800 (alternating/0.6)	1,466,835
max/jitter	247,574 (jitter/0.6)	247,953
max/missing	243,414 (missing/0.6)	243,414
max/bursty	199,730 (bursty/0.6)	199,882
max/noise	148,177 (noise/0.6)	147,811
worst/periodic	236,631 (periodic/1)	236,631
worst/alternating	1,466,800 (alternating/1)	1,466,835
worst/jitter	6,242,089 (jitter/1)	6,434,483
worst/missing	1,174,373 (missing/1)	1,174,373
worst/bursty	199,730 (bursty/1)	199,882
worst/noise	148,177 (noise/1)	147,811

Mutant evidence (run on scratch copies of this branch, full package suite):

Mutant in periodic.go	Work guard	Other tests red
Seed tolerance reintroduced: the refinement range uses tol(seed) instead of periodRange	red: fixture/alternating: 73761 steps in total, less than measured 104271 -25%: … lost search coverage …	TestPeriodicRelativeJitterRefinesWithTheCandidateTolerance, TestRefinedCandidateReplacesOnlyWhenItWouldSignal
periodRange memo removed (memoRange always recomputes)	green, as designed: visited gaps are unchanged	TestRangeMemoIsNeverStale: 0 memo hits of 200000 queries

With the old one-sided guard, the first mutant would have passed the guard: its steps fall from 104,271 to 73,761.

2. BenchmarkBytesPerStream (bench_test.go)

Before: the whole measurement sat inside for it := 0; it < b.N; it++, with the timer stopped for the full iteration. The measurement is a new Detector, 4 × n events and two forced GCs. Because the timed work was about 0 ns, the framework kept raising b.N towards 10⁹ and ran that many full measurements.

Now:

• Each sub-benchmark takes one heap measurement (bytesPerStream(n, periodic)) outside the b.N loop, with the timer stopped, and reuses it when the framework calls the sub-benchmark again with a larger b.N.
• The b.N loop is empty, and B/stream is reported with b.ReportMetric as before.
• ns/op is meaningless for this benchmark, and the comment says so.
• b.Loop() is not used, because the module's go.mod targets Go 1.22.

Command, on unchanged master 18a13264 vs this branch	before (master)	after (this branch)
go test -run '^$' -bench . -count=1 ./...	never finished: stopped by hand after 13 min 35 s, still in the first sub-benchmark BytesPerStream/streams=1000/periodic=false (-timeout 10m does not cover benchmarks)	63.5 s, 43 benchmark lines, exit 0
go test -run '^$' -bench . -benchtime=1x -count=1 ./...	4.0 s (5 s wall)	3.9 s (4 s wall)

Load during the runs: load average 3.4–3.9 on 12 cores.

B/stream is the same measurement as before, now taken exactly once per sub-benchmark:

streams	periodic	master (-benchtime=1x)	this branch (full run)
1,000	false	760.7	760.7
1,000	true	1345	1345
10,000	false	747.9	747.9
10,000	true	1332	1332
100,000	false	739.0	739.0
100,000	true	1323	1323

Test results (local, go1.26.0 darwin/arm64, in internal/anomaly)

• gofmt -l . lists nothing, and go vet ./... is clean.
• go test -race -count=1 ./...: ok (12.2 s).
• go test -run TestPeriodicSearchWorkIsBounded -count=3 -v ./...: 3/3 pass, and all 18 rule/shape totals are identical in all three runs (the counters are deterministic).
• go test -run '^$' -bench . -benchtime=1x -count=1 ./...: completes in 3.9 s.
• go test -run '^$' -bench . -count=1 ./...: completes in 63.5 s.

Scope

• Changes only internal/anomaly/periodic_bench_test.go and internal/anomaly/bench_test.go, in one commit each.
• No production code, golden, config, CI or deploy changes.

🤖 Generated with Claude Code