test: fold duplicated test/benchmark scaffolding into shared helpers #673

Test and benchmark scaffolding has been copied per package, next to shared helpers that already
exist. There are ~151 struct types in _test.go files. Recent PRs (#671, #672) each added another
heapSampler, inProcessS3, golden digest and backend matrix. This issue tracks folding them into
shared helpers. Survey at f30d3d4 (main + #672).

Existing shared helpers — extend these, don't invent new homes

• backend/backendtest:
  • Run conformance, AtomicConditionalPut, WithCapacity, WithDeferred.
  • It is used by internal package block tests, so it must never import block, engine or
    recordengine.
• backend/faultbackend: Wrap, Rule, Gate, Ops, Count. It forwards no optional
  capabilities.
• internal/mergestream/mergestreamtest:
  • a per-key read-counting Backend with FailWriteAt;
  • a cross-engine Run(t, open) suite through a Store adapter. This is the model for sharing
    suites across engines without an import cycle.
• internal/obs/obstest, internal/vfs/vfstest, internal/reproduce.

Plan (ranked: most duplication, lowest risk first)

• 1. Same-package merges (no boundary risk) — landed in test: fold duplicated test/benchmark scaffolding into shared helpers (#673, items 1-4) #674
  • root:
    • nodeCPUCorpus/headCorpus → parametrised;
    • isPartKey/goldenIsPartKey;
    • reportIngestMetrics/goldenReportPoints;
    • flushDefault(tb, s, signal) for ~55 inline lookup*Engine("default") + flush sites;
    • hand-rolled itoa → strconv.Itoa (also in engine, encoding/chunk and pool tests).
  • engine_test:
    • liveHeap/heapAlloc;
    • generalise flushParts (engine/decode_bench_test.go) for the 5 external "series × samples
      × parts → flush" corpus copies.
  • recordengine: extract the duplicated BenchmarkMergeCompact* loop body.
  • backend/s3: faultStore embeds fakeStore.
  • SetBytes literals *16 → engine.SampleBytes where the package can import it.
• 2. backendtest leaf helpers: — landed in test: fold duplicated test/benchmark scaffolding into shared helpers (#673, items 1-4) #674
  • WithoutCapabilities (5 copies);
  • StreamingMemory (2);
  • per-key read Counting (4, absorbing mergestreamtest.Backend via a type alias);
  • a capability-preserving byte counter (2);
  • Digest(tb, b, prefixes) for golden part hashing (2 identical + similar), keeping the golden
    format byte-identical;
  • PartDirs/IsPartObject (4).
• 3. backend/s3/s3test: Client/Backend/Shared/Case replace 5 in-process — landed in test: fold duplicated test/benchmark scaffolding into shared helpers (#673, items 1-4) #674
  go-faster/fs setups.
  • It always wraps backendtest.AtomicConditionalPut and closes idle connections.
  • That fixes a latent conditional-put race: storage_s3_test.go and both merge golden tests skip
    the wrapper today.
• 4. backendtest.Matrix + s3test.Case: one backend × window table for the two identical — landed in test: fold duplicated test/benchmark scaffolding into shared helpers (#673, items 1-4) #674
  merge golden matrices and the 3 benchmark backend tables.
• 5. Migrate ~25 hand-rolled fault/count wrappers to faultbackend. — test(faultbackend): replace hand-rolled fault and count backends (#673, item 5) #676
  • 13 of them are identical engine/recordengine twins: rejectReads, failReads, rejectWrites,
    rejectIndexWrites, countCAS, sizeKeyWrite, gateWrites/gateReads, invariantCAS,
    syncedCommits, readCounter, …
  • Needs first:
    • Op.Bytes;
    • Rule.Lose (a conditional write that loses without an error);
    • a gate mode that holds every match (today Gate is Times:1);
    • an After hook.
  • Medium risk: gate and error-identity semantics.
• 6. internal/heaptest: Live, Allocated, BytesPerOp, and FileSampler. — landed in test(heaptest): share heap and allocation measurement (#673, item 6) #675
  • FileSampler embeds *file.File, so capabilities still forward and the measurement doesn't
    change.
  • Also AssertFlat, with the floored growth ratio and an absolute bound.
  • About 12 sites. Keep every test's tuned floor, ratio and !race tag.
• 7. cluster/etcd/etcdtest: 4 embedded-etcd copies, every one getting the root's — etcdtest in test(etcdtest): share the embedded etcd and bind it to port 0 (#673, #509) #678 (also removes cluster/etcd tests are flaky in CI: port reuse race in freeAddr, and a timing-sensitive rejoin assertion #509's port race by binding port 0); fetchtest dropped: routing ~150 call sites through it was net +44 lines, all churn and no dedup
  port-collision retry. query/fetch/fetchtest.Collect: 9+ fetch-and-drain copies.
• 8. internal/enginetest: cross-engine suites for the ~52 test files mirrored between — test(enginetest): run the twin engine tests as one cross-engine suite (#673, item 8) #679; the fetcher-type unification it surfaced landed as refactor(bucketindex): declare the repair seam types once #681 (refactor: declare PartFetcher/FetchResult/RepairStats once instead of per engine #680)
  engine/ and recordengine/ (repair, hole, wants, replica…), through an adapter the way
  mergestreamtest does it. It is blocked on PartFetcher/FetchResult being distinct types per
  engine, unless the adapter bridges them.

Found along the way

recordengine's partsBytes doc says it sums on-disk size. It sums sizeBytes, which in the
record engine is decoded bytes (introspect.go), while in the metrics engine it is disk bytes. So
the two engines' merge benchmark MB/s are not comparable. Fix the doc in item 1, and state the unit
in both benchmarks.

Rules for every PR here

• Behaviour-preserving: no test's assertion, tuned threshold or measured quantity changes, and the
  golden files stay byte-identical.
• A shared helper that forwards or strips backend capabilities must say which, because several
  tests measure exactly that.
• Shared packages must not create import cycles: no engine or recordengine imports from anything
  those packages' internal tests use.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KcLq1G8KpWBvM53JaQb1tS