test(provisioner): real-backend gRPC round-trips for Mongo/Queue/Storage (Wave 4)

[mastermanas805]

What

Closes the remaining cells of the provisioner gRPC × backend matrix (INTEGRATION-COVERAGE-PLAN §2.3 / Wave 4). Postgres + Redis server-layer round-trips already shipped (#45, #46); this PR adds the real-backend Provision → assert artifact → Deprovision → assert-gone lifecycle for the three remaining backends — Mongo, Queue (NATS), Storage (MinIO/S3) — all driven through the genuine gRPC handlers (ProvisionResource/DeprovisionResource/GetStorageBytes/RegradeResource), not fakes.

Inventory (before → after)

RPC × Backend	Before	After
Provision/Deprovision/Regrade × postgres	EXISTS (#45/#46)	EXISTS
Provision/Deprovision × redis	EXISTS (#45/#46)	EXISTS
Provision/StorageBytes/Deprovision × mongo	GAP (fakes only)	EXISTS — real createUser/dropUser/dropDatabase + dbStats
Provision/Deprovision × queue (NATS)	GAP	EXISTS — real NATS health-check provision, no-op deprovision
GetStorageBytes × storage (MinIO)	GAP	EXISTS — real PutObject → measured bytes → delete → 0
GetStorageBytes × queue (=0), Regrade skip × mongo/queue	untested	EXISTS

Provisioner RPC surface = ProvisionResource, DeprovisionResource, GetStorageBytes, RegradeResource (proto provisioner/v1). Storage Provision/Deprovision are intentionally API-side; the provisioner only meters storage.

Tests added (internal/server/server_live_roundtrip_mqs_test.go)

• Mongo: provision creates usr_/db_ on real MongoDB; GetStorageBytes reads real dbStats (>0 after seeding a 4KB doc); Deprovision runs real dropUser/dropDatabase (truehomie DROP-incident class); 2nd Deprovision is a clean idempotent no-op; Regrade(mongo) asserts the documented skip path.
• Queue (NATS): provision passes the real :8222/healthz monitor check + returns nats:// URL + subject prefix; GetStorageBytes(queue)=0 (message-metered); Deprovision is the shared-backend no-op, idempotent; Regrade skip.
• Storage (MinIO): object-walk accounting — empty prefix=0, after a real PutObject=exact byte count (8192), after delete=0.

All env-gated: skip cleanly under go test -short (the deploy gate); run for real when backend env is present.

CI (.github/workflows/coverage.yml)

Added a NATS service container (JetStream + monitor :8222) and a MinIO docker run step (services can't pass server /data), plus env wiring TEST_NATS_HOST / TEST_MINIO_* / CUSTOMER_MONGO_AUTH_URL so the new tests execute. Mongo was already provided.

Integration-only coverage (touched package)

internal/server with all backends wired: 99.2% of statements. The five exercised handlers (provisionMongo, provisionQueue, GetStorageBytes, DeprovisionResource, RegradeResource) are all 100%. Method: go test ./internal/server/ -coverpkg=./internal/server/... -coverprofile=… -count=1 per the PLAN's mechanism C. The ≥80 integration floor is computed after subtracting the lines in the new INTEGRATION-COVERAGE-EXCLUSIONS.md (only k8s dedicated-backend boot wiring + cmd/ entrypoints — genuinely unreachable without a live kube-apiserver / forking the binary).

Verification

Ran all 8 server round-trip tests against real mongo/nats/minio/postgres/redis containers locally — all PASS. Failing-then-passing confirmed (mutated the storage expected-byte literal and the mongo >0 assert → both FAIL as expected, restored → PASS). Gate green: make gate (build + vet + go test ./... -short).

Bug found

None. The destroy/regrade paths for all three backends behaved correctly and idempotently end-to-end (relevant to the truehomie DROP-incident class — no partial-cleanup leak observed).

🤖 Generated with Claude Code