test(provisioner): real-backend integration round-trips for Deprovision/Regrade DDL (toward 100% integration)

[mastermanas805]

What

Env-gated real-backend integration tests that drive the gRPC server layer's Provision → Regrade → Deprovision lifecycle against a real Postgres and a real Redis, through the actual RPC handlers (server.ProvisionResource / RegradeResource / DeprovisionResource) — not injected fakes.

Why (truehomie-db DROP incident class, 2026-06-03)

Every existing server_test.go / server_coverage_test.go test injects a fake backend, so the real DROP DATABASE / DROP USER / ALTER ROLE DDL had never executed through the gRPC handler path (breaker wrapping, tier→connLimit routing, mapError, response shaping, idempotent re-deprovision). 98.9% statement coverage from mocks did not prove the destroy/regrade DDL is correct end-to-end — exactly the regression class where a real DROP path hides.

Tests (all t.Skip cleanly when the backend URL is unset → CI without a backend stays green)

• TestServer_Postgres_Provision_Regrade_Deprovision_LiveRoundTrip — asserts db_/usr_ created, role CONNECTION LIMIT actually ALTERed (cross-checked against pg_roles.rolconnlimit), db_/usr_ DROPped, and a second Deprovision is a clean idempotent no-op (the fix(bugbash 2026-05-21): Provisioner P1 wave-1 — Redis k8s sizingForTier yearly + plus tiers #9 DROP IF EXISTS fix).
• TestServer_Postgres_Reprovision_AfterDeprovision_LiveRoundTrip — re-provisions the same token after teardown; guards a partial-DROP leak that would block reuse.
• TestServer_Redis_Provision_Deprovision_LiveRoundTrip — ACL user created, ACL user + namespace keys reaped, idempotent second Deprovision. (Redis LocalBackend has no Regrade — only the k8s backend implements redis.Regrader.)

Coverage delta (integration-only, per §1.4 mechanism — -coverpkg=./...)

Measured from the server test package (the gRPC layer):

Metric	Before	After
server test pkg vs ./...	18.9%	24.8% (+5.9pp)
backend/postgres/local.go Provision (from gRPC layer)	0.0%	68.3%
backend/postgres/local.go Deprovision	0.0%	62.5%
backend/postgres/local.go Regrade	0.0%	73.3%
backend/redis/local.go Provision	0.0%	70.0%
backend/redis/local.go Deprovision	0.0%	83.3%

The backend DDL methods, as exercised FROM the gRPC handlers, go from 0% to 60–83% — the real destroy/regrade path is now integration-covered, not just unit-covered in isolation.

Env-blocked / out of scope (this increment)

• mongo / queue (NATS) / storage backend round-trips remain GAP — no local TEST_MONGO_URL/TEST_NATS_URL. Postgres + Redis prioritized per the bounded scope.

Gate

• go build ./... && go vet ./... && go test ./... -short — green (exit 0, no FAIL).
• gofmt (via $(go env GOROOT)/bin/gofmt) + golangci-lint — clean (0 issues).
• Verified locally: all three tests PASS against local Postgres (localhost:5432) + Redis (localhost:6379), and SKIP cleanly with backends unset.

🤖 Generated with Claude Code