diff --git a/sw-block/design/test/v3-phase-15-g15a-privileged-qa-test-instruction.md b/sw-block/design/test/v3-phase-15-g15a-privileged-qa-test-instruction.md
new file mode 100644
index 000000000..097b353de
--- /dev/null
+++ b/sw-block/design/test/v3-phase-15-g15a-privileged-qa-test-instruction.md
@@ -0,0 +1,284 @@
+# V3 Phase 15 G15a Privileged QA Test Instruction
+
+Date: 2026-05-03
+Status: privileged-path verified on m01 at `p15-g15a/csi-static-mvp@ac49adb`
+Scope: real `iscsiadm` discovery+login + `mkfs.ext4` + `mount` + bind-mount + filesystem write/read byte-equal + clean teardown, all driven through `cmd/blockcsi` NodeStage / NodePublish against real product daemons
+Companion: `v3-phase-15-g15a-qa-test-instruction.md` (non-privileged L2 control-plane verification)
+
+## Headline
+
+At `p15-g15a/csi-static-mvp@ac49adb`, on Linux host m01 (192.168.1.181, Ubuntu
+6.17 kernel, open-iscsi, sudo NOPW), G15a's `cmd/blockcsi` binary drives the
+full privileged data-plane sequence end-to-end: `iscsiadm` discovery, login,
+device-by-IQN wait, `mkfs.ext4`, mount-to-staging, bind-mount-to-target, 4KB
+random byte-equal write/read through the mounted device, `NodeUnpublish`
+(unmount target), `NodeUnstage` (unmount staging + iscsiadm logout),
+`ControllerUnpublish`, with no dangling iSCSI sessions remaining.
+
+## Environment
+
+- **Repo**: `seaweed_block`
+- **Branch**: `p15-g15a/csi-static-mvp`
+- **Minimum commit**: `ac49adb`
+- **Host**: m01 (192.168.1.181) — Linux 6.17, open-iscsi, sudo NOPW for `testdev`
+- **SSH**: `ssh -i C:/work/dev_server/testdev_key testdev@192.168.1.181`
+- **SMB harness path**: `V:\share\g15a-priv\` (Windows) = `/mnt/smb/work/share/g15a-priv/` (Linux)
+- **Go**: 1.26.2 (project default)
+- **Privileges required**: root (test invokes `sudo` and runs the test binary as root so `iscsiadm` and `mount` work)
+- **Why m01 specifically**: the only host in our test fleet with kernel iSCSI initiator + sudo NOPW + clean baseline; G15a-K8s (real kubelet PVC) is forward-carry to a dedicated K8s test cluster
+
+## Pre-conditions
+
+Before running, verify on m01:
+
+```bash
+ssh -i C:/work/dev_server/testdev_key testdev@192.168.1.181 '
+  which iscsiadm mkfs.ext4 sudo go &&
+  go version &&
+  sudo -n true && echo SUDO_NOPW_OK &&
+  sudo iscsiadm -m session 2>&1 | head -3
+'
+```
+
+Expected:
+
+- `iscsiadm`, `mkfs.ext4`, `sudo`, `go` all resolved
+- Go ≥ 1.24
+- `SUDO_NOPW_OK` printed
+- `iscsiadm -m session` reports `No active sessions` (or no session for `iqn.2026-05.io.seaweedfs:g15a-priv-v1`)
+
+If any baseline iSCSI session exists for the test IQN, log out before running:
+
+```bash
+sudo iscsiadm -m node -T iqn.2026-05.io.seaweedfs:g15a-priv-v1 --logout || true
+```
+
+## Run command
+
+### One-shot via the harness runner with persistent artifacts
+
+The harness at `V:\share\g15a-priv\` contains:
+
+- `privileged_test.go` — the Go test runner (single file, ~250 LOC). Reads `G15A_ARTIFACT_DIR` env var; if set, all per-daemon logs land there instead of `t.TempDir()`.
+- `run-g15a-privileged.sh` — bash wrapper that builds binaries, stages the test, compiles, runs under sudo, prints baseline + post-run iSCSI sessions, and stages logs to a per-run artifact directory.
+
+Ship a fresh source tree to m01 and invoke with a persistent run directory:
+
+```bash
+RUN_ID="$(date -u +%Y%m%dT%H%M%SZ)"
+
+# from your dev box (Windows or Linux)
+git -C /path/to/seaweed_block archive --format=tar HEAD | \
+  ssh -i C:/work/dev_server/testdev_key testdev@192.168.1.181 \
+    'rm -rf /tmp/g15a-priv/src && mkdir -p /tmp/g15a-priv/src && tar x -C /tmp/g15a-priv/src'
+
+# trigger the run on m01 with persistent artifact dir
+ssh -i C:/work/dev_server/testdev_key testdev@192.168.1.181 "
+  cp /mnt/smb/work/share/g15a-priv/privileged_test.go /tmp/g15a-priv/src/cmd/blockcsi/privileged_test.go
+  cd /tmp/g15a-priv/src
+  go build -o /tmp/g15a-priv/bin/blockmaster ./cmd/blockmaster
+  go build -o /tmp/g15a-priv/bin/blockvolume ./cmd/blockvolume
+  go build -o /tmp/g15a-priv/bin/blockcsi ./cmd/blockcsi
+  go test -c -o /tmp/g15a-priv/blockcsi-priv.test ./cmd/blockcsi
+  mkdir -p /mnt/smb/work/share/g15a-priv/runs/$RUN_ID
+  sudo G15A_PRIVILEGED=1 G15A_BIN_DIR=/tmp/g15a-priv/bin \
+       G15A_ARTIFACT_DIR=/mnt/smb/work/share/g15a-priv/runs/$RUN_ID \
+       /tmp/g15a-priv/blockcsi-priv.test \
+       -test.run TestG15aPriv -test.v -test.timeout=180s \
+       2>&1 | tee /mnt/smb/work/share/g15a-priv/runs/$RUN_ID/test-stdout.log
+"
+echo "Artifacts at: V:\\share\\g15a-priv\\runs\\$RUN_ID\\"
+```
+
+Expected wall clock: ~10 s build + ~6 s test = ~16 s total. Artifacts persist
+on the SMB share for inspection without an additional SSH round-trip.
+
+> **Important**: the `G15A_ARTIFACT_DIR` env var is what makes logs survive a
+> green test. Without it, `privileged_test.go` falls back to `t.TempDir()` and
+> Go cleans up on success, leaving only the test stdout phase markers — no
+> per-daemon logs to inspect.
+
+### Artifact directory layout
+
+After a green run, `V:\share\g15a-priv\runs\<RUN_ID>\` contains:
+
+| File / dir | Size (typical) | Purpose |
+|---|---|---|
+| `test-stdout.log` | ~1 KB | Test framework output: 6 phase markers + PASS line. Read first to see the timeline. |
+| `blockmaster.log` | ~270 B | Master daemon stderr. Quiet — startup, lock acquired, listen address. |
+| `blockvolume-primary.log` | **~20 KB** | **Most informative log.** Contains: `g7-debug` adapter R/S/H state transitions, real iSCSI session lifecycle (kernel initiator IQN visible), every SCSI WRITE/SYNCHRONIZE_CACHE during `mkfs.ext4`, replication barrier waits with r2 at each sync, the test payload write at `lba=0`, and session invalidate on cleanup. |
+| `blockvolume-replica.log` | ~270 B | r2 observes r1 is primary, records supersede, doesn't apply locally. |
+| `topology.yaml` | ~130 B | The 2-slot RF=2 spec the master loaded. |
+| `staging/` `store-primary/` `store-replica/` `authority-store/` | empty post-cleanup | Mount points + durable roots. Empty after `NodeUnstage` — that emptiness is itself evidence cleanup worked. |
+
+**Currently missing**: `blockcsi.log` (the CSI driver's own iscsiadm/mount call log) is **not** in the persisted dir. The current harness writes it to `t.TempDir()` directly, bypassing `G15A_ARTIFACT_DIR`. Forward-carry: extend `startG15aCSI` helper to honor `art` parameter (3-line change). Workaround until then: wrap with `strace -f -e trace=execve` to capture iscsiadm/mkfs/mount syscalls.
+
+### What the dev agent can verify directly from the logs
+
+From `blockvolume-primary.log`:
+
+1. **Real kernel iSCSI initiator attached** (look for):
+   ```
+   session: FullFeature initiator="iqn.2004-10.com.ubuntu:01:..." target="iqn.2026-05.io.seaweedfs:g15a-priv-v1" type="Normal"
+   ```
+   (Discovery session before this gets reset by peer — normal two-phase iscsiadm behavior, not a bug.)
+
+2. **mkfs.ext4 generated ~40 SCSI WRITEs** for filesystem metadata (search `iscsi: SCSI WRITE handler lba=`). The test payload write is the one with `lba=0 transferLen=1 dataOut=4096`.
+
+3. **Every WRITE replicated to r2 with barrier handshake** — for the test payload at `lsn=42`:
+   ```
+   replication: OnLocalWrite volume=v1 lba=0 lsn=42 peers=1
+   replication: ship ok peer=r2 lba=0 lsn=42
+   replication: barrier wait peer=r2 addr=... targetLSN=42 epoch=1
+   replication: barrier ack peer=r2 targetLSN=42 achievedLSN=42
+   ```
+   Proves byte-equal payload was both durable on primary AND replicated to r2 before the read.
+
+4. **Cleanup observed** at the end:
+   ```
+   executor: invalidate session 1 for r2: replica peer closed
+   ```
+
+### Manual / debug invocation (no harness wrapper)
+
+If you want to iterate on the test logic without re-shipping source, the test
+file lives at `/mnt/smb/work/share/g15a-priv/privileged_test.go`. Edit it
+directly via SMB, then on m01 run the inner block of the SSH heredoc above.
+
+## Phase-by-phase scenario
+
+The single test `TestG15aPriv_NodeStageBindWriteReadCleanup` drives 6 phases.
+Each phase logs its own success line so you can see in `-test.v` output exactly
+where a failure happened.
+
+### Phase 1 — ControllerPublishVolume returns matching frontend fact
+
+- Calls `ControllerPublishVolume(VolumeId=v1, NodeId=node-a)` against `blockcsi`.
+- Polls up to 15 s; passes when `publish_context["iscsiAddr"]` and `["iqn"]` match the values seeded into `blockvolume`'s iSCSI listen flags.
+- Log: `Phase 1 OK: ControllerPublish returned iscsiAddr=... iqn=...`
+
+### Phase 2 — NodeStageVolume drives iscsiadm + mkfs + mount
+
+- Calls `NodeStageVolume` with the `publish_context` from Phase 1.
+- Internally `cmd/blockcsi` invokes `iscsiadm -m discovery` then `iscsiadm -m node ... --login`, waits for `/dev/disk/by-path/*<iqn>*` to appear, then `mkfs.ext4` and `mount` to the staging path.
+- Verified by `mountpoint -q <staging>` (with `/proc/mounts` fallback).
+- Log: `Phase 2 OK: NodeStage completed (iscsiadm discovery+login + mkfs + mount)` + `Phase 2 verified: <staging> is a mountpoint`
+
+### Phase 3 — NodePublishVolume bind-mounts staging → target
+
+- Calls `NodePublishVolume` to bind-mount staging into the target path.
+- Verified by `mountpoint -q <target>`.
+- Log: `Phase 3 OK: NodePublish completed (bind mount staging → target)`
+
+### Phase 4 — Byte-equal write/read through the CSI mount
+
+- Generates 4096 random bytes via `crypto/rand`.
+- Writes to `<target>/g15a-priv-byteequal.bin`.
+- Calls `sync` to flush page cache.
+- Reads back, asserts byte-equal.
+- Log: `Phase 4 OK: 4096 bytes round-tripped byte-equal through CSI-mounted iSCSI device`
+
+### Phase 5 — Explicit cleanup chain
+
+- `NodeUnpublishVolume` → unmount target.
+- `NodeUnstageVolume` → unmount staging + `iscsiadm --logout`.
+- `ControllerUnpublishVolume` → soft-tolerated (G15a static MVP may return Unimplemented; the test logs and continues).
+- Log: `Phase 5 OK: NodeUnpublish + NodeUnstage clean`
+
+### Phase 6 — No dangling iSCSI session
+
+- Runs `iscsiadm -m session`, asserts the test IQN is **not** present in output.
+- Log: `Phase 6 OK: no dangling iSCSI session for iqn.2026-05.io.seaweedfs:g15a-priv-v1`
+
+## QA evidence pins
+
+### Pin 1 — first verification (ephemeral logs, t.TempDir())
+
+```
+Tested by: QA
+Date: 2026-05-03
+Tree: p15-g15a/csi-static-mvp@ac49adb
+Host: m01 (192.168.1.181), Linux 6.17, Go 1.26.2 linux/amd64
+Result: PASS in 5.69s
+Baseline iSCSI sessions: none
+Post-run iSCSI sessions: none
+Logs: ephemeral (t.TempDir, cleaned by Go on success)
+```
+
+### Pin 2 — re-run with persistent artifact dir
+
+```
+Tested by: QA
+Date: 2026-05-03
+Tree: p15-g15a/csi-static-mvp@ac49adb (same source)
+Host: m01 (192.168.1.181), Linux 6.17, Go 1.26.2 linux/amd64
+Result: PASS in 5.93s
+Baseline iSCSI sessions: none
+Post-run iSCSI sessions: none
+Run ID: 20260503T153623Z
+Artifact dir: V:\share\g15a-priv\runs\20260503T153623Z\
+Persisted: test-stdout.log, blockmaster.log, blockvolume-primary.log (19 KB),
+           blockvolume-replica.log, topology.yaml
+Missing: blockcsi.log (forward-carry harness fix)
+```
+
+Phases all green in both runs:
+
+1. ControllerPublish returned matching iscsiAddr + iqn
+2. NodeStage (iscsiadm discovery+login + mkfs.ext4 + mount) + mountpoint verified
+3. NodePublish (bind mount staging → target) + mountpoint verified
+4. 4096 random bytes round-tripped byte-equal through CSI-mounted iSCSI device
+5. NodeUnpublish + NodeUnstage clean (ControllerUnpublish acceptable Unimplemented at static MVP)
+6. No dangling iSCSI session for the test IQN
+
+```
+Test file: V:\share\g15a-priv\privileged_test.go
+Runner script: V:\share\g15a-priv\run-g15a-privileged.sh
+```
+
+Re-run on a different commit will produce a fresh evidence pin; the two pins
+above are canonical for the privileged-path verification at `ac49adb`.
+
+## Non-claims (this instruction does not validate)
+
+1. **K8s integration** — no real `kubelet` / PVC / Pod / scheduler involved. Real kubelet integration is forward-carry to a dedicated K8s test cluster.
+2. **Multi-node attach** — single node (m01), single ControllerPublish/NodeStage cycle. Multi-attach (RWX) is not tested.
+3. **Partition table / filesystem variety** — only `ext4` direct on the device. No XFS, no LVM, no GPT partitioning.
+4. **Block volume mode** — only mount-mode (filesystem) tested. Raw block-mode CSI capability is not exercised.
+5. **Concurrent write/read across reconnect** — single sequential write+sync+read. Concurrent or post-restart write/read is not covered.
+6. **Failover during stage** — primary is not killed mid-NodeStage. G8 covers primary-kill data continuity at the storage layer; CSI-layer failover under workload is forward-carry.
+7. **Persistence across plugin restart** — test starts and stops the `cmd/blockcsi` plugin once; "restart plugin then unstage stale volume" is not exercised.
+8. **CSI snapshot / clone / resize** — G15a static MVP does not advertise these capabilities; not relevant.
+9. **Performance / IOPS / throughput** — single 4KB I/O, no perf claim. G21 owns SLO.
+10. **CHAP / mutual auth / TLS** — beta-defer until G16.
+
+## Forward-carry
+
+| Item | Owner | Where it lands |
+|---|---|---|
+| **Plugin restart + stale-volume unstage**: kill `cmd/blockcsi` mid-staging, restart, verify `NodeUnstageVolume` cleans up | sw + QA on m01 | G15a-6 followup or G15a-K8s |
+| **Multi-node attach (single-writer enforcement)**: second node attempts NodeStage same volume, expect rejection | QA on m01 + a second Linux host | G15a-RWO test (separate instruction) |
+| **Block-mode capability**: NodeStage with `Block{}` access type instead of `Mount{}` | sw must wire through, QA verifies | G15a block-mode followup |
+| **Real K8s kubelet driving the same flow**: PVC → Pod schedules → kubelet calls CSI → workload writes/reads | QA | G15a-K8s on dedicated cluster |
+| **Failure injection during NodeStage** (e.g., iscsiadm timeout, mkfs failure) | sw + QA | G15a-fault followup |
+| **Post-G8 failover under live mount**: kill primary while file is being written, verify CSI re-discovery / device re-attach behavior | sw + QA | G15a × G8 cross-gate test |
+| **`blockcsi.log` capture in artifact dir**: extend `startG15aCSI` helper to honor `art` (3-line change) so the CSI driver's own iscsiadm/mount call log lands alongside daemon logs | QA harness | trivial follow-up; bundle with first edit to `privileged_test.go` |
+
+## Cleanup / re-run safety
+
+The test always registers `t.Cleanup` for both `NodeUnpublishVolume` and
+`NodeUnstageVolume` so even on test failure mid-flight, the iSCSI session and
+mounts get torn down. If a test panic somehow leaves residue, manual cleanup:
+
+```bash
+sudo umount /tmp/TestG15aPriv*/001/target 2>/dev/null || true
+sudo umount /tmp/TestG15aPriv*/001/staging 2>/dev/null || true
+sudo iscsiadm -m node -T iqn.2026-05.io.seaweedfs:g15a-priv-v1 --logout 2>/dev/null || true
+sudo iscsiadm -m node -T iqn.2026-05.io.seaweedfs:g15a-priv-v1 -o delete 2>/dev/null || true
+```
+
+After cleanup, `sudo iscsiadm -m session` should report `No active sessions`.
+
+## Close sentence
+
+> *G15a privileged QA proves the V3 `cmd/blockcsi` binary correctly drives the full Linux block storage stack — kernel iSCSI initiator, ext4 format, kernel mount, bind-mount, filesystem I/O — against real V3 product daemons, with byte-equal data correctness and clean teardown leaving no kernel-level residue.*
diff --git a/sw-block/design/test/v3-phase-15-g15a-qa-test-instruction.md b/sw-block/design/test/v3-phase-15-g15a-qa-test-instruction.md
new file mode 100644
index 000000000..2594515b1
--- /dev/null
+++ b/sw-block/design/test/v3-phase-15-g15a-qa-test-instruction.md
@@ -0,0 +1,171 @@
+# V3 Phase 15 G15a QA Test Instruction
+
+Date: 2026-05-03
+Status: non-privileged L2 close-candidate on `p15-g15a/csi-static-mvp@ac49adb`
+Scope: blockcsi static MVP control-plane path through real product daemons; non-privileged only
+
+## Headline
+
+At `p15-g15a/csi-static-mvp@ac49adb`, G15a proves a real `blockcsi` driver can
+serve CSI Identity, look up assigned frontend facts from a real `blockmaster`
+status path, and return a populated `publish_context` from `ControllerPublishVolume`
+that matches the actual `iscsiAddr` and `iqn` exposed by a real `blockvolume`
+iSCSI target.
+
+This instruction covers what runs on plain CI / Windows / unprivileged Linux.
+Privileged steps (real `iscsiadm` discovery+login, `mkfs`, `mount`, file write/read)
+require an authenticated Linux test host and are out of scope for this instruction;
+see §Non-claims and §Forward-carry.
+
+## Environment
+
+- Repo: `seaweed_block`
+- Branch: `p15-g15a/csi-static-mvp`
+- Minimum commit: `ac49adb`
+- Fidelity: subprocess L2 with real `cmd/blockmaster` + real `cmd/blockvolume` (×2 for r1/r2) + real `cmd/blockcsi`
+- Platform: any Go-capable host (Windows, macOS, plain Linux without iscsiadm); no `root` / privileged container required
+- Go version: project default (CGO_ENABLED=1 on Windows; no `-race` flag on Windows per project policy)
+
+## Full Regression Command
+
+```bash
+go test ./core/csi ./cmd/blockcsi ./core/host/volume ./core/host/master ./core/authority ./cmd/blockmaster ./cmd/blockvolume -count=1
+```
+
+Expected: all 7 packages PASS.
+
+Wall-clock guidance on a typical dev box:
+
+| Package | Approx. time |
+|---|---|
+| `core/csi` | < 0.1 s |
+| `cmd/blockcsi` | ~25-30 s (includes L2 build + subprocess spawn) |
+| `core/host/volume` | < 0.1 s |
+| `core/host/master` | < 0.2 s |
+| `core/authority` | ~7 s |
+| `cmd/blockmaster` | < 0.1 s |
+| `cmd/blockvolume` | ~120-150 s (includes G8 + G9 subprocess tests) |
+| **Total** | ~3 min |
+
+## Focused G15a L2 Command
+
+```bash
+go test ./cmd/blockcsi -run TestG15a_BlockCSIControllerPublishUsesMasterFrontendFact -count=1 -v
+```
+
+Expected: PASS in ~20-25 s on a local dev box. Includes building all three binaries
+(`blockmaster`, `blockvolume`, `blockcsi`) in a temp dir, spawning them, polling
+`ControllerPublishVolume` for up to 15 s until the published `iscsiAddr`/`iqn`
+match the values the test seeded into the volume.
+
+## Scenario Checklist
+
+Each scenario lists the backing test so dev agents can re-run a single scenario
+in isolation when iterating.
+
+### CSI service-layer scenarios (component, fast)
+
+1. **Identity service handles nil requests safely.**
+   - `core/csi/identity_endpoint_test.go::TestIdentity_NilRequests`
+   - `core/csi/identity_endpoint_test.go::TestParseEndpoint`
+
+2. **CSI package does not import authority and does not construct AssignmentFacts.**
+   Structural guard from `v3-phase-15-control-plane-evolution.md §7.1`.
+   - `core/csi/boundary_guard_test.go::TestG15a_CSI_DoesNotImportAuthorityOrConstructAssignmentFacts`
+
+3. **`ControllerPublishVolume` returns iSCSI publish_context built from a verified target fact.**
+   - `core/csi/controller_test.go::TestControllerPublish_ReturnsISCSIPublishContextFromTargetFact`
+
+4. **`ControllerPublishVolume` fails closed when no verified target exists.**
+   - `core/csi/controller_test.go::TestControllerPublish_FailsClosedWithoutVerifiedTarget`
+
+5. **`ControllerPublishVolume` fails closed when the target has no frontend fact attached.**
+   - `core/csi/controller_test.go::TestControllerPublish_FailsClosedWhenTargetHasNoFrontendFact`
+
+6. **`ControllerPublishVolume` propagates backend lookup errors as `Internal`.**
+   - `core/csi/controller_test.go::TestControllerPublish_PropagatesLookupErrorsAsInternal`
+
+7. **Controller capabilities do not advertise dynamic provisioning at G15a.**
+   - `core/csi/controller_test.go::TestControllerCapabilities_G15aDoesNotAdvertiseDynamicProvisioning`
+
+8. **`ValidateVolumeCapabilities` confirms an existing target without minting state.**
+   - `core/csi/controller_test.go::TestValidateVolumeCapabilities_ConfirmsExistingTarget`
+
+9. **`NodeStage` honors `publish_context` over `volume_context` when both are present.**
+   - `core/csi/node_test.go::TestNodeStage_UsesPublishContextBeforeVolumeContext`
+
+### Master backend lookup scenarios (component, fast)
+
+10. **Master backend maps iSCSI status frontend into `TargetFact`.**
+    - `core/csi/master_backend_test.go::TestControlStatusLookup_MapsISCSIStatusFrontend`
+
+11. **Master backend maps NVMe status frontend into `TargetFact`.**
+    - `core/csi/master_backend_test.go::TestControlStatusLookup_MapsNVMeStatusFrontend`
+
+12. **Master backend fails closed when no frontend has been assigned.**
+    - `core/csi/master_backend_test.go::TestControlStatusLookup_FailClosedWithoutAssignedFrontend`
+
+### L2 subprocess scenario (the headline)
+
+13. **Real `blockmaster` + `blockvolume` (×2) + `blockcsi` ⇒ `ControllerPublish` returns the volume's actual frontend fact.**
+    - `cmd/blockcsi/main_test.go::TestG15a_BlockCSIControllerPublishUsesMasterFrontendFact`
+    - This is the headline integration scenario. The test:
+      - builds all three binaries in `t.TempDir()`,
+      - launches `blockmaster`,
+      - launches `blockvolume` r1 with seeded iSCSI listen address and IQN,
+      - launches `blockvolume` r2,
+      - launches `blockcsi --master <addr>`,
+      - polls `ControllerPublishVolume(VolumeId="v1", NodeId="node-a")` for up to 15 s,
+      - asserts `publish_context["iscsiAddr"]` equals r1's seeded `iscsiAddr` and `publish_context["iqn"]` equals r1's seeded IQN.
+    - Pass condition: published context matches seeded values within the deadline.
+    - Skip condition: `-short` flag.
+
+### Binary smoke scenario
+
+14. **`blockcsi` binary starts and serves Identity.**
+    - `cmd/blockcsi/main_test.go::TestBlockCSI_BinaryStartsAndServesIdentity`
+
+## Non-claims (this instruction does not validate)
+
+The following are explicitly **out of scope** for this instruction. Do not assume
+they pass on the basis of this checklist:
+
+1. **Real `iscsiadm` discovery / login** — requires Linux + initiator + privileged execution.
+2. **Real `mkfs.<fs>` against the attached block device** — requires kernel block device + privileged execution.
+3. **Real `mount` / bind-mount into a target path** — requires privileged execution + mount namespace.
+4. **Filesystem write/read of test files through the mounted CSI volume** — requires the full privileged path above.
+5. **`NodeUnstage` / `NodeUnpublish` cleanup against real OS state** — requires the full privileged path above.
+6. **Kubernetes CSI driver registration / kubelet integration** — requires a real K8s node.
+7. **Multiple concurrent `ControllerPublish` calls / N-node `NodeStage` fan-out** — single happy path only.
+8. **Snapshot / clone / resize CSI verbs** — G15a does not advertise these capabilities.
+9. **CSI security / RBAC / TLS** — beta-defer until G16.
+10. **Performance / SLO** — G21.
+
+## Forward-carry (next gate / next environment)
+
+| Item | Owner | Where it lands |
+|---|---|---|
+| **G15a-5 M01 privileged**: real `iscsiadm` discovery+login, `mkfs`, `mount`, file write/read byte-equal, `NodeUnstage`/`NodeUnpublish` cleanup | sw + QA on M01 | next G15a sub-step; uses the same `cmd/blockcsi` binary, just on a privileged Linux host |
+| **CSI dynamic provisioning** (`CreateVolume`/`DeleteVolume`) | sw | post-G15a / G15b mini-plan |
+| **CSI snapshot/clone/resize** | sw | G10 / G11 / G15b separate gates |
+| **CSI security (TLS / token)** | sw | G16 |
+| **Kubernetes integration test** (real kubelet, real PVC, real pod) | QA | G15a-K8s sub-step on M01 or dedicated K8s test cluster |
+
+## Close sentence
+
+> *G15a non-privileged L2 proves the V3 product loop produces an assignment, blockvolume exposes a real iSCSI frontend fact, and `blockcsi`'s `ControllerPublishVolume` correctly retrieves attach information from master status — all through real product daemons, with no privileged OS operations.*
+
+## QA evidence on this instruction (one-time pin)
+
+```
+Tested by: QA
+Date: 2026-05-03
+Tree: p15-g15a/csi-static-mvp@ac49adb
+Focused L2: TestG15a_BlockCSIControllerPublishUsesMasterFrontendFact PASS 22.13s
+Full 7-pkg regression: all PASS
+Worktree: C:/work/seaweed_block_g9c
+Platform: Windows 11 Pro (Go default, CGO_ENABLED=1, no -race per project policy)
+```
+
+Re-run on a different host or commit will produce a fresh evidence pin; this
+block is the canonical pin for the close-candidate state at `ac49adb`.
diff --git a/sw-block/design/v3-phase-15-g15a-close-report.md b/sw-block/design/v3-phase-15-g15a-close-report.md
new file mode 100644
index 000000000..59314ea1b
--- /dev/null
+++ b/sw-block/design/v3-phase-15-g15a-close-report.md
@@ -0,0 +1,157 @@
+# V3 Phase 15 — G15a CSI Static MVP Close Report
+
+**Date**: 2026-05-03
+**Status**: CLOSED — static/pre-provisioned CSI iSCSI path verified
+**Code branch**: `p15-g15a/csi-static-mvp`
+**Close anchor**: `seaweed_block@ac49adb`
+**Docs/test evidence**:
+- `sw-block/design/test/v3-phase-15-g15a-qa-test-instruction.md`
+- `sw-block/design/test/v3-phase-15-g15a-privileged-qa-test-instruction.md`
+- `V:\share\g15a-priv\runs\20260503T153623Z\`
+
+---
+
+## §1 Close Claim
+
+G15a proves the static/pre-provisioned V3 CSI iSCSI path:
+
+```
+cluster-spec / product-loop assignment
+  -> blockmaster authority assignment
+  -> blockvolume iSCSI frontend target fact
+  -> blockcsi ControllerPublish publish_context
+  -> blockcsi NodeStage real Linux iSCSI login + mkfs + mount
+  -> blockcsi NodePublish bind mount
+  -> filesystem write/read byte-equal
+  -> clean unpublish/unstage/logout
+```
+
+CSI remains a consumer of V3 product facts. It does not mint authority, create volumes, own storage, or derive target identity when a frontend target fact exists.
+
+---
+
+## §2 Code Delivered
+
+| Commit | Summary |
+|---|---|
+| `7b10413` | Added `core/csi` skeleton, CSI Identity/Controller/Node services, boundary guards, and iSCSI NodeStage/Publish mechanism. |
+| `4e150ca` | Added frontend target fact propagation: `blockvolume` heartbeat -> master observation/status -> CSI `ControlStatusLookup`. |
+| `94ff9cf` | Added `cmd/blockcsi` static driver binary with Identity/Controller/Node registration. |
+| `ac49adb` | Added non-privileged L2 smoke proving `ControllerPublish` returns real `blockvolume` frontend facts through `blockcsi`. |
+
+---
+
+## §3 Evidence
+
+### §3.1 Non-Privileged L2
+
+Tree: `p15-g15a/csi-static-mvp@ac49adb`
+
+Commands:
+
+```powershell
+go test ./cmd/blockcsi -run TestG15a_BlockCSIControllerPublishUsesMasterFrontendFact -count=1 -v
+go test ./core/csi ./cmd/blockcsi ./core/host/volume ./core/host/master ./core/authority ./cmd/blockmaster ./cmd/blockvolume -count=1
+```
+
+Result: PASS.
+
+What this proves:
+- `cmd/blockcsi` starts and serves CSI Identity.
+- `blockmaster`, `blockvolume`, and `blockcsi` run as real subprocesses.
+- `blockvolume` reports the iSCSI frontend target fact.
+- `blockmaster QueryVolumeStatus` exposes the assigned replica's frontend fact.
+- `ControllerPublishVolume` returns `publish_context["iscsiAddr"]` and `["iqn"]` matching the real `blockvolume` target.
+
+### §3.2 Privileged M01 L3
+
+Run ID: `20260503T153623Z`
+
+Artifact directory:
+- Windows: `V:\share\g15a-priv\runs\20260503T153623Z\`
+- m01: `/mnt/smb/work/share/g15a-priv/runs/20260503T153623Z/`
+
+Key files:
+- `test-stdout.log`
+- `blockmaster.log`
+- `blockvolume-primary.log`
+- `blockvolume-replica.log`
+- `topology.yaml`
+
+Result: PASS.
+
+The privileged test verified six phases:
+
+| Phase | Verified operation |
+|---|---|
+| 1 | `ControllerPublishVolume` returned expected `iscsiAddr` and `iqn`. |
+| 2 | `NodeStageVolume` completed real `iscsiadm` discovery/login, `mkfs.ext4`, and mount; staging path was a mountpoint. |
+| 3 | `NodePublishVolume` bind-mounted staging to target; target path was a mountpoint. |
+| 4 | 4096 random bytes wrote and read back byte-equal through the CSI-mounted iSCSI device. |
+| 5 | `NodeUnpublishVolume` and `NodeUnstageVolume` completed cleanly. |
+| 6 | No dangling iSCSI session remained for the test IQN. |
+
+Additional log evidence from `blockvolume-primary.log`:
+- Real Linux kernel initiator reached iSCSI `FullFeature` for target `iqn.2026-05.io.seaweedfs:g15a-priv-v1`.
+- `mkfs.ext4` generated real SCSI `WRITE` and `SYNCHRONIZE_CACHE` traffic.
+- The final payload write path included `OnLocalWrite`, `ship ok peer=r2`, and replication barrier acknowledgements.
+
+Known artifact note: `blockcsi.log` was not persisted in this run because the harness helper wrote it to an ephemeral temp directory. This is a harness artifact gap, not a product-path gap; `test-stdout.log` and `blockvolume-primary.log` contain the relevant externally visible CSI and iSCSI/data-plane evidence.
+
+---
+
+## §4 Boundary Guards Preserved
+
+G15a preserves the V3 control-plane boundaries:
+
+- CSI package does not import `core/authority` or `core/adapter`.
+- CSI package does not construct `AssignmentAsk`, `AssignmentFact`, or `adapter.AssignmentInfo`.
+- CSI `ControllerPublish` reads frontend target facts through a narrow `PublishTargetLookup`.
+- `blockmaster` status exposes frontend target facts as read-only observation/status data.
+- `blockvolume` remains the owner of iSCSI/NVMe target identity and serving.
+
+This matches the G15a unification rule: CSI is attach/mount orchestration, not storage ownership or authority.
+
+---
+
+## §5 Non-Claims
+
+G15a does not claim:
+
+- Kubernetes PV/PVC/Pod integration.
+- Dynamic CSI `CreateVolume`.
+- CSI snapshot, clone, or expansion.
+- NVMe CSI attach path.
+- Multi-node RWO enforcement.
+- CHAP, mutual auth, TLS, or routable frontend security.
+- Pod remount after primary kill.
+- Failover while a filesystem is mounted.
+- Plugin restart cleanup after crash.
+- Performance, throughput, or soak behavior.
+
+---
+
+## §6 Forward-Carry
+
+| Item | Target |
+|---|---|
+| Persist `blockcsi.log` into privileged artifact dirs. | Harness polish; not a G15a blocker. |
+| Kubernetes static PV/PVC + pod write/read using `cmd/blockcsi`. | G15b. |
+| Plugin restart + stale-volume `NodeUnstage` cleanup. | G15b/G15c follow-up. |
+| Multi-node attach and RWO enforcement. | G15c. |
+| Failover under live mounted filesystem. | G15c/G16, after G8 failover semantics are exercised through CSI. |
+| NVMe NodeStage/NodePublish path. | G15d. |
+| Dynamic `CreateVolume` from CSI into desired-volume / placement allocation. | G15e or later, after product API ratification. |
+
+---
+
+## §7 Close Decision
+
+G15a is closed on `seaweed_block@ac49adb`.
+
+The static CSI iSCSI path is proven at both:
+
+- non-privileged L2 control-plane fidelity, and
+- privileged m01 Linux data-plane fidelity.
+
+Next gate should be G15b: Kubernetes static PV/PVC/pod integration using the same `cmd/blockcsi` binary and the same static/pre-provisioned volume model.