Phoebe: make reconcile index EXPLAIN test robust (perf=soft, structural=hard) + dedupe

[hhuuggoo]

Test-quality cleanup pass on the rating tests. The battery on PR #19 found zero production-code issues — every finding was about the EXPLAIN/index integration tests being fragile or redundant. This fixes those so the suite is robust.

Stacking note: intended to stack on yaml-rater-fixes-5, but that branch was merged (PR #19) and deleted while this work was in flight. PR #19 merged into yaml-rater-fixes-4 (merge commit 301e9f6), which now contains all of fixes-5's content. This PR is therefore based on and targets yaml-rater-fixes-4 — no fixes-5 content is lost (the trees were identical).

Contracts

NO production or contract change. Test-quality only. All three changes are in test files (internal/rating/store_integration_test.go, internal/rating/rater_test.go); no .go production source, schema, migration, or behavior was touched. The reconcile/rating semantics, log severities, exit codes, and the rated_usage contract are all unchanged. All pre-existing tests still pass.

Fixes

FIX 1 — TestIntegration_ReconcileDeleteUsesWindowStartIndex → …CanUseWindowStartIndex: separate the structural gate from the perf signal.
The test conflated two claims: (a) the window_start index can serve the reconcile's window-only predicate (a structural/correctness invariant), and (b) the planner prefers it at default cost over the auth-leading composite (a performance assertion coupled to Postgres's cost model, fragile across PG majors / GUCs). A perf property whose failure mode is a slow reconcile — not a wrong bill — must not be a fragile correctness-style gate.

• Removed the vacuous Seq Scan on rated_usage ru guard. Proven vacuous in the live plan: the priced WHERE-false CTE collapses to a One-Time Filter: false, so rated_usage's only access is the DELETE target; dropping the window_start index falls back to the composite index scan, never a literal Seq Scan — so the guard could never fire independently of the index assertion above it.
• The default-cost check is now discriminating (asserts the index by NAME via the index-scan node, using rated_usage_window_start_ix, not a bare substring the composite would also match) but a SOFT signal — t.Logf, never t.Fatalf.
• The HARD gate (t.Fatalf) is the enable_seqscan=off force-the-index check: the structural proof the index can serve the predicate at all.
• Renamed so the name claims only what is hard-asserted — the index is usable, not that the planner picks it.

FIX 2 — dedupe the explainPlan helper (fix-plus-cleanup).
TestIntegration_RatingInstantIndexServesScan had an inline EXPLAIN-row-scanning loop identical in intent to the explainPlan helper PR #19 introduced — same single-column scan, same plan-row concatenation, same rows.Err() handling, same plain EXPLAIN (no ANALYZE/FORMAT options). Confirmed no subtle difference, then swapped it to explainPlan and deleted the duplicated loop.

FIX 3 — correct an inaccurate docstring on TestRater_RoutineReconcileDeleteLogsError.
The docstring claimed it was RED before the fix because "testLogger discards INFO" — but this test uses captureLogger(), which buffers both the INFO and ERROR streams (it does not discard INFO; testLogger() is a different logger). Corrected to the actual mechanism: the assertion is on which severity the line fired at; the old always-INFO code left the captured ERROR stream empty, failing errBuf.Len()==0. (Test names/docstrings are claims — fixed per that rule.)

Gate

go build ./..., go vet ./... (+ -tags=integration), golangci-lint v1.64.8 (+ --build-tags=integration), gofmt -l . — all clean. go test -race ./... and PHOEBE_TEST_DATABASE_URL=… go test -tags=integration -race ./... (live PG 16) — all pass, including the e2e suite and the rebuilt index EXPLAIN tests (ran, not skipped). Verified the renamed test's soft-log fires and the live plan confirms the removed Seq-Scan guard was vacuous (One-Time Filter: false).

🤖 Generated with Claude Code

[hhuuggoo]

Landed on main via the squashed rating merge (3b22908). Closing the stack.