diff --git a/portal/docs/FLEET-DESIGN.md b/portal/docs/FLEET-DESIGN.md
new file mode 100644
index 0000000..984e0d8
--- /dev/null
+++ b/portal/docs/FLEET-DESIGN.md
@@ -0,0 +1,344 @@
+# Fleet ingest — design (locked)
+
+Status: **approved**, ready for Phase 13 implementation.
+
+This is the design reference for the Admin / Fleet aggregation feature. Read this before touching any code in Phases 13–15.
+
+---
+
+## 1. Goals & non-goals
+
+**In scope (v1):**
+- Mirror every customer's raw `monitoring.PowerMeasurements`, `Sites`, and `Devices` into a central DB on an Admin stack, with `CustomerId` attached.
+- Admin operator sees fleet-wide and per-customer dashboards.
+- Customer stacks keep working if Admin is unreachable; pushes resume when it comes back.
+- Same portal binary in two modes (`RunMode=Client` / `Admin`), config-selected.
+
+**Not in scope (v1):**
+- Cross-customer cost / tariff aggregation (Admin shows kWh + power; per-customer instance still owns billing).
+- Branding / user sync (Admin has its own brand + users — separate fleets of humans).
+- Bidirectional sync (Admin → Customer commands, OTA, etc.).
+- Multi-region or sharded central DB.
+
+---
+
+## 2. Topology
+
+```
+[Traefik public]
+    ├─ abc0001.portal.example.com  →  Client stack (portal + timescale + grafana)
+    ├─ abc0002.portal.example.com  →  Client stack
+    ├─ ...
+    └─ admin.portal.example.com    →  Admin stack (portal + timescale + grafana)
+
+[Client stack abc0001] --batches POST--> [Admin stack /api/fleet/ingest]
+[Client stack abc0002] --batches POST--> (central TimescaleDB)
+[Client stack ...]     --batches POST-->
+```
+
+Admin stack reuses `docker-compose.prod.yml` and the same image, with:
+- `COMPOSE_PROJECT_NAME=admin`
+- `CUSTOMER_HOST=admin.portal.example.com`
+- `RunMode=Admin`
+- Its own Postgres volume (the central fleet DB)
+
+The whole feature is opt-in. With no Admin stack deployed, customer stacks set `FleetIngest__Enabled=false` and work exactly as today.
+
+---
+
+## 3. `RunMode` mechanics
+
+`Application.RunMode` is `Client` (default) or `Admin`. Single binary; Program.cs branches:
+
+| Concern | Client | Admin |
+|---|---|---|
+| DbContext target | local Postgres | central Postgres |
+| Endpoints `/api/sites/*`, `/api/measurements/*`, `/api/ingest/measurements`, `/api/admin/sites/*` | mapped | **not mapped** |
+| Endpoints `/api/fleet/*`, `/api/admin/customers/*` | not mapped | mapped |
+| `FleetPushService` (BackgroundService) | started if `FleetIngest__Enabled=true` | not registered |
+| `FleetIngestService` | not registered | registered |
+| EF migration set | `app` + `monitoring` | `app` + `fleet` |
+| Identity / Branding / Users / Settings | same | same |
+| Settings → Rates tab | shown | hidden |
+| Sidebar nav | Dashboard / Dashboards / Sites / Settings | Dashboard / Dashboards / **Customers** / Settings |
+
+**Startup guards:**
+- Reject `RunMode=Admin` if `Database:ConnectionString` is empty (no auto-provision for Admin — there is no obvious local default for a fleet DB).
+- Reject `RunMode=Client + FleetIngest__Enabled=true` if `FleetIngest__Url` or `FleetIngest__Token` is empty.
+- Reject the wrong mode against an existing DB by sniffing for a marker table:
+  - `RunMode=Client` against a DB containing `fleet.Customers` → fatal.
+  - `RunMode=Admin` against a DB containing `monitoring.PowerMeasurements` → fatal.
+- Stops the "I pointed dev at the prod DB" disaster.
+
+---
+
+## 4. Schemas
+
+### 4.1 Central DB (Admin) — `fleet` schema
+
+**`fleet.Customers`** — registry
+| Column | Type | Notes |
+|---|---|---|
+| Id | uuid PK | Admin-generated, stable |
+| Code | varchar(50) UNIQUE | e.g. `ABC0001` — human handle (preserve case for display, lowercase for compose) |
+| Name | varchar(200) | |
+| TokenHash | varchar(64) UNIQUE indexed | SHA-256 hex of the push token |
+| TokenIssuedAt | timestamptz | |
+| TokenRotatedAt | timestamptz nullable | |
+| IsActive | bool | If false, ingest returns 401 |
+| FirstSeenAt, LastSeenAt | timestamptz nullable | When their first / most-recent push landed |
+| CreatedAt | timestamptz | |
+
+**`fleet.Sites`** — mirrors customer-side, identity preserved
+| Column | Type | Notes |
+|---|---|---|
+| Id | uuid | Customer-side Site UUID, preserved |
+| CustomerId | uuid FK→Customers | |
+| Name, Address, IsActive | … | |
+| LocalMunicipalityId | int nullable | Opaque on Admin |
+| ReceivedAt | timestamptz | When Admin upserted this row |
+| **PK** | (CustomerId, Id) | |
+
+**`fleet.Devices`** — same pattern
+| Column | Type | Notes |
+|---|---|---|
+| Id | uuid | Preserved |
+| CustomerId | uuid FK | |
+| SiteId | uuid | (CustomerId, SiteId) FK→Sites composite |
+| Name, ExternalId, Description, IsActive | … | |
+| ReceivedAt | timestamptz | |
+| **PK** | (CustomerId, Id) | |
+
+**`fleet.PowerMeasurements`** — hypertable, the big one
+| Column | Type | Notes |
+|---|---|---|
+| Time | timestamptz | Hypertable partition |
+| CustomerId | uuid | |
+| DeviceId | uuid | (CustomerId, DeviceId) FK→Devices |
+| ActivePowerKw, ReactivePowerKvar, ApparentPowerKva, PowerFactor, VoltageV, FrequencyHz, EnergyImportedKwh, EnergyExportedKwh | double | |
+| Source | varchar(50) nullable | |
+| **PK** | (Time, CustomerId, DeviceId) | Time must be in any unique constraint on a hypertable |
+| **Indexes** | (CustomerId, Time DESC); (CustomerId, DeviceId, Time DESC) | |
+
+**`fleet.IngestEvents`** — observability
+| Column | Type | Notes |
+|---|---|---|
+| Id | uuid PK | |
+| CustomerId | uuid FK | |
+| ReceivedAt | timestamptz | |
+| BatchType | varchar(20) | `sites` / `devices` / `measurements` |
+| RowsAccepted, RowsRejected | int | |
+| BatchBytes | int | |
+| ClientHwm | varchar(50) | Cursor the client claims this batch covers |
+| TimeSpread | interval nullable | `max(Time) - min(Time)` in the batch — visible burst back-fills |
+| Error | varchar(500) nullable | |
+
+### 4.2 Client DB additions
+
+One new table: **`app.FleetPushState`**
+| Column | Type | Notes |
+|---|---|---|
+| ResourceType | varchar(20) PK | `sites` / `devices` / `measurements` |
+| LastCursor | timestamptz nullable | For measurements: max(`ReceivedAt`) pushed. For sites/devices: max(`UpdatedAt`) |
+| LastSyncedAt | timestamptz nullable | |
+| LastError | varchar(500) nullable | |
+| ConsecutiveFailures | int default 0 | Drives exponential backoff |
+
+Plus: **add `ReceivedAt timestamptz` (default `NOW()`) to `monitoring.PowerMeasurements`**, indexed on `(ReceivedAt)`. See §5.1 for the rationale.
+
+### 4.3 Migration assembly split
+
+Two `DbContext` classes — `ClientDbContext` (existing `AppDbContext`, renamed) and `AdminDbContext`. Each owns its migrations folder. Only the one matching `RunMode` is registered with DI. Ugly-but-reliable wins over single-context-with-tooling-gymnastics.
+
+---
+
+## 5. Push pipeline
+
+### 5.1 Late arrivals — push by `ReceivedAt`, not `Time`
+
+The firmware (ESP32 / TTGO T-Call) **buffers samples offline** and replays them when connectivity returns. Late arrivals are a designed-for case, not an edge.
+
+If we pushed by `Time > LastCursor`, a device offline for 6 hours would have its back-fill skipped (its samples' `Time` is in the past).
+
+Fix: every row carries `ReceivedAt` assigned by the local DB on insert. Push selects `WHERE ReceivedAt > LastCursor[measurements]`. Back-fills are picked up on the next push tick after the firmware replays.
+
+`Time` stays as-is for queries; nothing else changes about how the data is used.
+
+### 5.2 Burst handling
+
+When a firmware replay drops thousands of rows in one local insert, the next push tick sees a huge backlog. To avoid one customer monopolising network / Admin:
+- Per request: up to `FleetIngest__BatchSize` rows (default 5000) **or** `FleetIngest__BatchMaxBytes` (default 1 MB), whichever hits first.
+- Per tick: at most 3 successful batches per resource type, then yield.
+- Result: a 30k-row back-fill drains over ~5 minutes (10 ticks @ 60s) without starving sites/devices syncing.
+- Observable: `fleet.IngestEvents.TimeSpread` shows the `max(Time) - min(Time)` per batch; a wave of back-fill is obvious in Admin logs without grepping.
+
+### 5.3 Cadence
+
+- Default push interval: 60 s.
+- Configurable: `FleetIngest__IntervalSeconds`.
+
+### 5.4 Order of operations per tick
+
+1. **Sites**: `UpdatedAt > LastCursor[sites]` → POST batch (full upsert, all rows; small data).
+2. **Devices**: same.
+3. **Measurements**: `ReceivedAt > LastCursor[measurements]` ordered by `ReceivedAt` → POST batches until drained or limit.
+
+Sites and devices first so Admin's measurement FK insert doesn't reject rows for unknown devices.
+
+### 5.5 Endpoint contract
+
+```
+POST /api/fleet/ingest
+Headers:
+  X-Customer-Token: <opaque 32-byte hex>
+  X-Batch-Type:     sites | devices | measurements
+  X-Push-Cursor:    <ISO8601> (highest ReceivedAt/UpdatedAt in this batch)
+  Content-Type:     application/json
+Body:
+  JSON array of rows matching the type
+Response 200:
+  { accepted: int, rejected: int, errors: [{ row: int, reason: string }] }
+Response 401: unknown / inactive / wrong token
+Response 413: batch too large
+Response 429: rate-limited
+Response 5xx: transient — retry
+```
+
+CustomerId is derived from the token server-side. Clients never send it. Prevents one customer from forging rows for another.
+
+### 5.6 Failure handling
+
+| Failure | Client behavior |
+|---|---|
+| Network timeout / 5xx | Exponential backoff (1m → 2m → 5m → 10m, cap 30m); `ConsecutiveFailures++` |
+| 401 | Stop trying for this resource; log loudly; surface in **Settings → App config** ("Admin link down") |
+| 200 with `rejected > 0` on measurements | Re-push sites/devices on next tick before retrying measurements |
+| 413 | Halve `BatchSize` for this resource and retry |
+| 429 | Honor `Retry-After`; otherwise back off 1 minute |
+
+Push runs on its own DI scope and its own DbContext. If the push DB query hangs, only push is blocked — never user-facing API.
+
+---
+
+## 6. Auth — token lifecycle
+
+- Token = 32 random bytes, hex-encoded (64 chars). Generated server-side via `RandomNumberGenerator`.
+- Stored as **SHA-256 hex hash**, UNIQUE indexed → ingest endpoint does a single O(log N) indexed lookup. Bcrypt is the wrong tool for a high-throughput, high-entropy machine token.
+- Shown **once** in the Admin UI at creation/rotation. Admin only stores the hash. Lost token → rotate.
+- Rotation: **immediate cutover** — new value replaces old hash. Customer ops updates `.env` + restarts. Brief push gap acceptable; can revisit with dual-token grace window later.
+- `Customers.IsActive=false` → all that customer's pushes get 401 until reactivated.
+
+---
+
+## 7. Continuous aggregates + retention
+
+### 7.1 Aggregates (created post-migration by `FleetTimescaleBootstrapper`)
+
+- `fleet_hourly_per_device` — `time_bucket('1 hour', Time)`, `CustomerId`, `DeviceId` → avg/max/min `ActivePowerKw`, `max(EnergyImported) - min(EnergyImported)` as `KwhDelta`, sample count.
+- `fleet_daily_per_customer` — `time_bucket('1 day', Time)`, `CustomerId` → totals across all that customer's devices.
+- **Realtime** (`materialized_only = false`) so late back-fills appear in CA queries immediately, with the unmaterialized tail served live from the hypertable until next refresh.
+- Refresh policy:
+  - `fleet_hourly_per_device`: every 5 min; `start_offset = 30 days`, `end_offset = 1 hour`.
+  - `fleet_daily_per_customer`: every 1 hour; `start_offset = 365 days`, `end_offset = 1 day`.
+- Wide `start_offset` means firmware back-fills within those windows get materialized on the next refresh tick. TimescaleDB's invalidation log means only chunks that actually changed get re-processed.
+
+### 7.2 Retention vs compression
+
+**Retention policy: NONE.** All data kept indefinitely for historical / trend analysis.
+
+**Compression policy** (this is what makes "forever" cheap):
+- `fleet.PowerMeasurements`: compress chunks older than **7 days**.
+- Compression settings: `segmentby = (CustomerId, DeviceId)`, `orderby = Time DESC`.
+- Typical ratio for power-monitoring data: **5–10×**. Point queries by `(CustomerId, DeviceId, Time range)` stay fast because of the segmentby clustering.
+- Bootstrapper enables compression and adds the policy idempotently.
+
+CA tables are tiny by comparison; no compression needed there.
+
+---
+
+## 8. Admin UI surface
+
+| Page | What |
+|---|---|
+| Dashboard | Fleet headline — customer count, active customers (push in last hour), aggregate live active power, total kWh today, last-24h lag chart |
+| Customers | Table — Code / Name / Last seen / Sites / Devices / Today kWh / Status / Actions (rotate token, disable, view) |
+| Customer detail | Sites + devices for the selected customer + Grafana embed scoped to `customer_id` |
+| Dashboards | Grafana embed — `fleet-overview` default, `customer-drilldown` parameterized |
+| Settings → Branding | Admin's own brand |
+| Settings → Users | Admin operator accounts |
+| Settings → Grafana | Read-only info card |
+| Settings → App config | Read-only config snapshot (RunMode visible at the top) |
+
+Hidden in Admin mode: **Sites** top-level (no local sites), **Settings → Rates** tab (no local tariffs).
+
+---
+
+## 9. Observability
+
+Two questions a real operator asks, two answers:
+
+1. **"Is customer X pushing?"** — Customers page shows `LastSeenAt`, `ConsecutiveFailures`, last batch size, recent burst spread.
+2. **"Why did this batch not land?"** — `SELECT * FROM fleet.IngestEvents WHERE CustomerId = … ORDER BY ReceivedAt DESC LIMIT 50;`
+
+Plus Serilog:
+- Client: `[INF] Pushed N measurements to fleet ingest (cursor=…) accepted=N rejected=0`
+- Admin: `[INF] Accepted N measurements from {Code} ({CustomerId}), rejected 0`
+
+---
+
+## 10. Per-stack config
+
+**Customer `.env` additions:**
+```
+RunMode=Client                                          # default — can omit
+FleetIngest__Enabled=true
+FleetIngest__Url=https://admin.portal.example.com/api/fleet/ingest
+FleetIngest__Token=<32-byte hex token from Admin Customers page>
+FleetIngest__IntervalSeconds=60
+FleetIngest__BatchSize=5000
+FleetIngest__BatchMaxBytes=1048576
+```
+
+`FleetIngest__Enabled=false` → push service doesn't start; customer runs as today, no data leaves.
+
+**Admin `.env`:**
+```
+RunMode=Admin
+COMPOSE_PROJECT_NAME=admin
+CUSTOMER_HOST=admin.portal.example.com
+Database__ConnectionString=Host=timescaledb;...        # central fleet DB (separate volume)
+# All other vars same as Client compose.
+```
+
+---
+
+## 11. Open seams (deferred, with the obvious extension paths noted)
+
+| Seam | Where v2 picks it up |
+|---|---|
+| Tariff sync + cross-customer cost | New `fleet.Tariffs` table, push tariff change events from customer, derived cost in `fleet_daily_per_customer`. |
+| Per-customer Postgres RLS for multi-Admin-user setups | Add `current_setting('app.customer_filter')`-based RLS policies on `fleet.*` tables; Admin role to customer-scope mapping in `IdentityRole.CustomerId`. |
+| Bidirectional Admin → Customer commands | New WebSocket or long-poll channel on customer side; gated by mutual cert or a second token. |
+| Branding sync (for the "Admin sees customer's brand when drilling in" niceness) | Push branding row from customer; Admin renders the customer's brand on Customer-detail pages. |
+| Forward compatibility | Versioned URL `/v1/fleet/ingest`. Admin tolerates unknown JSON fields. Strictness on response shape only. |
+| Dual-token grace window for rotation | `Customers.PreviousTokenHash` column; ingest accepts either for 24h after rotation. |
+| Sharded Admin (10000+ customers) | Customer's `FleetIngest__Url` already supports any host — point different customers at different Admin instances; aggregate at a tier above with Promscale or similar if needed. |
+| Hard-delete / GDPR | Admin Customers page → "Decommission" action: `DELETE FROM fleet.* WHERE CustomerId = …` cascade. Logged. |
+
+---
+
+## 12. Phase split
+
+| Phase | What | End-state |
+|---|---|---|
+| **13** | `RunMode` plumbing + Fleet schema + `AdminDbContext` split + Customers registry + token issuance UI. No data movement. | Spin up an Admin stack, register a customer, see their token. Client stack picks up `RunMode=Client` + push config but doesn't yet push. |
+| **14** | `FleetPushService` (Client), `/api/fleet/ingest` + `FleetIngestService` (Admin), `IngestEvents`, `ReceivedAt` migration on Client, `FleetPushState`, continuous aggregates SQL, compression policy SQL. | Two local stacks: one Client pushes to one Admin. Rows visible in `fleet.PowerMeasurements`. Late-arrival firmware-replay scenario verified with fixtures. |
+| **15** | Admin UI (Dashboard, Customers list, Customer detail, Customers token rotation), Grafana provisioning of `fleet-overview.json` + `customer-drilldown.json`, README + OPERATIONS updates for Admin deployment + customer onboarding workflow. | Full feature usable end-to-end. |
+
+---
+
+## 13. Sanity checks (carried over from the design conversation)
+
+- **No central data without the central DB existing.** Whole feature is opt-in via `FleetIngest__Enabled`.
+- **Customer model is unchanged.** Every customer keeps full control of their own data. We copy *roll-ups* (and raw measurements for fidelity), not authority.
+- **One ingest endpoint, no fan-out.** N customers all POST to the same `/api/fleet/ingest`. Admin scales vertically first, shards horizontally later if needed (each customer's `FleetIngest__Url` can point at a specific Admin instance).