diff --git a/.claude/ui-overhaul.md b/.claude/ui-overhaul.md
new file mode 100644
index 0000000..c5caf17
--- /dev/null
+++ b/.claude/ui-overhaul.md
@@ -0,0 +1,300 @@
+# Hub Manager UI Overhaul: Follow-up Work
+
+The web client was redesigned (black + Cornell red, top tab bar, side sheets, three-step flash
+plan) in web-client `feat/ui-overhaul`. That change is **frontend only**: no cloud-services,
+rpi-hub-server or API contract changes. This file lists what the backend and contract need for
+the new UI to be fully supported, with STM32 flashing first, and the frontend shortcuts to
+delete once each backend change lands.
+
+Written for humans and agents. Each item says why it matters, what to change in which repo,
+and when it is done. Do them in the order of the table below unless there is a reason not to.
+
+## Rules (read first)
+
+- **The contract comes first.** Wire types live in cloud-services (`src/models.py`,
+  `src/protocol/bench_v1.py`). After changing them, run `python -m src.protocol.export` in
+  cloud-services to rewrite `contracts/openapi.json` (CI fails if it is stale), then
+  `npm run gen:types` in web-client. Never hand-edit `web-client/src/types/api.gen.ts`.
+- **Keep the mock in step.** `web-client/src/mock/` is typed from the contract. Update it in the
+  same PR so `npm run dev:mock` shows the new behavior and `npm run type-check` passes.
+- **Old hubs keep working.** Hubs are updated one at a time, so every new field is optional and
+  the UI must behave as it does today when a field is missing.
+- Branches are `feat/...` or `fix/...` with a PR into `main`. Merge submodule PRs first and the
+  hyperloop-gui PR (which bumps the submodule pointers) last, because merging it deploys.
+
+## What shipped (frontend only)
+
+| Area | Where in web-client |
+|---|---|
+| Design tokens (colors, fonts) | `src/index.css`, the `@theme` block |
+| Header, socket handlers, sheet host, toasts | `src/components/layout/AppShell.tsx` |
+| Pages: Hubs, Devices, Telemetry, Flash | `src/pages/*.tsx` |
+| Side sheets: hub, device, terminal, add streams, schema | `src/components/sheets/` |
+| SVG charts, merge, export | `src/components/telemetry/ChartPanel.tsx` |
+| Ports and connections per hub (shared) | `src/stores/deviceStore.ts` |
+| Open sheet, toasts | `src/stores/uiStore.ts` |
+
+### Frontend shortcuts to remove later
+
+The UI works around missing backend support in these places. Each points to the item that
+replaces it.
+
+| Shortcut | File | Replaced by |
+|---|---|---|
+| A pending flash is shown as in progress (animated bar, elapsed time) because hubs never report `running` | `src/pages/Flash.tsx` (`FlashProgress`) | 1 |
+| One "Compile and write" timeline row, since there are no stages | `src/pages/Flash.tsx` (`FlashProgress`) | 2 |
+| Which tool flashes a board (`arduino-cli` or `openocd`) is a hard-coded map | `src/config/boards.ts` (`FLASHERS`, `flasherFor`) | 3 |
+| The board list duplicates `rpi-hub-server/config/boards.yaml` | `src/config/boards.ts` (`BOARDS`) | 3 |
+| Idle connections are closed by the browser, only while someone has the Devices page open | `src/stores/deviceStore.ts` (`enforceInactivity`), "Quiet · closes in" hint in `src/pages/Devices.tsx` | 8 |
+| Device counts on Hubs cost two requests per hub every 30 s | `src/stores/deviceStore.ts` (`refresh`) | 7 |
+| CPU, memory and disk show "—" for up to 30 s after a page load | `src/pages/Hubs.tsx`, `src/components/sheets/HubSheet.tsx` | 6 |
+
+## Work list
+
+| # | Change | Repos | Contract change | Size |
+|---|---|---|---|---|
+| 1 | Hub reports `running` when a task starts | rpi-hub-server | no | S |
+| 2 | Flash stages and progress (`stage`, `progress`) | all three | yes | M |
+| 3 | Hub reports each board's flasher; cloud serves the board list | all three | yes | M |
+| 4 | Typed flash result and a "Show log" view | cloud-services, web-client | yes | S |
+| 5 | Cancel a running flash | all three | yes | M |
+| 6 | Latest health in REST (no empty meters on load) | cloud-services, web-client | yes | S |
+| 7 | Port and connection counts in the hub list | cloud-services, web-client | yes | S |
+| 8 | Close idle serial connections on the hub, not in the browser | rpi-hub-server, web-client | no | S |
+| 9 | Terminal and chart history after a reload | cloud-services, web-client | yes | M |
+| 10 | Custom chart schemas actually used for parsing | web-client (then cloud-services) | later | S |
+| 11 | Self-host Uber Move | web-client | no | S |
+| 12 | Frontend polish: code splitting, phone layout, tests | web-client | no | M |
+
+Items 1 to 5 complete STM32 flashing support. Items 6 to 12 are independent.
+
+## STM32 flashing: full support
+
+What already works end to end: the STM32F407G-DISC1 is detected by USB id (`boards.yaml`,
+`disco_f407vg`), `.ino` is compiled with arduino-cli and programmed with OpenOCD over the
+on-board ST-LINK (`write, verify, reset`), `.bin/.elf/.hex` images skip the compile, several
+ST-LINKs on one hub are told apart by serial number (`OpenOcdFlasher.adapter_serial`), and the
+Flash page accepts those formats and says "OpenOCD over ST-LINK". What is missing is
+visibility while it runs, control over a running flash, and one source of truth for boards.
+
+### 1. Report `running` when a task starts
+
+**Why:** `rpi-hub-server/src/bench/command_handler.py` reports a task when it is queued
+(`handle_command`) and when it finishes (`_task_worker`), never in between. `BaseTask.run()`
+sets `RUNNING` internally (`src/bench/tasks/base_task.py`) but nobody sends it. A 5 to 15 minute
+STM32 compile therefore looks queued until it is done. The mock backend sends `running`, which
+is why `npm run dev:mock` does not show the problem.
+
+**Change (rpi-hub-server):** in `_task_worker`, emit a status right after the task starts:
+create the run task, wait until `task.status == RUNNING` (or emit from inside `BaseTask.run()`
+through a callback), then call `self._report_task_status(task)`. No contract change: `running`
+is already a valid status.
+
+**Then (web-client):** nothing is required; `hubStore.updateTaskStatus` records `started_at` and
+the timeline marks the step active. Optionally restore distinct pending copy in
+`FlashProgress` ("Queued on <hub>") now that pending really means queued.
+
+**Done when:** flashing an `.ino` to the DISC1 shows "Compiling and flashing…" with the
+"Compile and write" row active and its start time filled in while the compile runs.
+
+### 2. Flash stages and progress
+
+**Why:** one bar and one row cannot tell "compiling for 6 minutes" from "stuck". The contract
+already has `TaskStatusBroadcast.progress` (0 to 100) and the UI already draws it (a determinate
+bar replaces the animated one as soon as `progress` is set), but no hub sends it.
+
+**Contract (cloud-services):** add an optional `stage` to the hub's task status message in
+`src/protocol/bench_v1.py` and to `TaskStatusBroadcast`:
+`"queued" | "compiling" | "writing" | "verifying" | "resetting"`. Pass it through unchanged like
+`progress`.
+
+**Hub (rpi-hub-server):**
+- `src/uplink/hub_agent.py` `send_task_status_update`: forward `stage` (it already forwards
+  `progress`).
+- Flash tasks report `stage` at each step and `progress` where the tool exposes it. Tool runs
+  need streaming output instead of waiting for exit (`run_tool` in
+  `src/bench/flashing/tools.py`): add a per-line callback.
+- OpenOCD's `program` command prints fixed markers on stderr: `** Programming Started **`,
+  `** Programming Finished **`, `** Verify Started **`, `** Verified OK **`,
+  `** Resetting Target **`. Map them to `writing`, `verifying`, `resetting`.
+- arduino-cli: `compiling` while `compile` runs (no percentage; keep `progress` null), then
+  `writing`. avrdude's upload prints `Writing | ####` progress bars, which can be turned into a
+  percentage.
+- Rate-limit progress messages (say one per second): they go over the cellular uplink too.
+
+**Web-client:** add `stage` to `Task` (`src/types/index.ts`) and store it in
+`hubStore.updateTaskStatus`. In `FlashProgress` (`src/pages/Flash.tsx`), replace the single
+"Compile and write" row with Compile (`.ino` only), Write, Verify (OpenOCD only) and Reset, each
+active or done from `stage`. Update `src/mock/server.ts` to step through the stages.
+
+**Done when:** an STM32 `.ino` flash walks Compile, Write, Verify, Reset in the timeline, and
+the bar fills during the write.
+
+### 3. One source of truth for boards and flashers
+
+**Why:** the web client hard-codes the board list (`BOARDS`) and which boards use OpenOCD
+(`FLASHERS`) in `src/config/boards.ts`, copied from `rpi-hub-server/config/boards.yaml`. Adding
+a board (for example another STM32 Discovery or Nucleo) means editing both, and they drift.
+
+**Hub (rpi-hub-server):** `Board.summary()` in `src/bench/boards.py` also returns `flasher` and
+`reset`, so every `PortInfo.board_profile` carries them. Send the full registry summary (id,
+name, fqbn, artifacts, flasher) in the handshake.
+
+**Contract (cloud-services):** add optional `flasher` (`"arduino-cli" | "openocd" | "none"`) and
+`reset` to `BoardProfileInfo`. Add `GET /api/boards`, the union of the registries the connected
+hubs reported, so the Flash page's "Board" dropdown lists what the hubs can actually flash.
+
+**Web-client:** `flasherFor()` reads `board.flasher` and falls back to the current map only for
+old hubs; the Board dropdown loads from `/api/boards` with `BOARDS` as the offline fallback.
+Delete both once every hub is updated.
+
+**Done when:** adding a board to `boards.yaml` on a hub makes it appear in the Flash page's
+Board dropdown, with the right "Flashed with …" line, without a web-client change.
+
+### 4. Typed flash result and "Show log"
+
+**Why:** flash results already carry the tool output: OpenOCD returns
+`{board_fqbn, flash_duration_ms, output, flasher}` with the last 2000 characters of its log.
+`TaskStatusBroadcast.result` is an untyped dictionary, so the UI does not show it. When a flash
+fails, the log is the first thing anyone needs.
+
+**Contract (cloud-services):** define `FlashResult` (`output`, `flash_duration_ms`, `flasher`,
+`board_fqbn`) and document it on `TaskStatusBroadcast.result` for `commandType == "flash"`.
+
+**Web-client:** under the timeline in `FlashProgress`, add a "Show log" disclosure with the
+output in the terminal style (mono, black background) and the duration.
+
+**Done when:** a failed STM32 flash shows the OpenOCD error text on the Flash page.
+
+### 5. Cancel a running flash
+
+**Why:** `BaseTask.cancel()` exists on the hub, but there is no cloud endpoint or command for
+it. A wrong board choice or a hung ST-LINK can only be waited out (the GUI gives up after
+17 minutes, `DEFAULT_TIMEOUT_MS.flash` in `src/services/commandService.ts`).
+
+**Contract and cloud (cloud-services):** `POST /api/hubs/{hub_id}/tasks/{task_id}/cancel`
+(operators only), sending a `cancel` command to the hub.
+
+**Hub (rpi-hub-server):** handle `cancel` in `command_handler.py`: call `task.cancel()` on a
+running task (kill the arduino-cli or OpenOCD process), drop a queued one, and report
+`cancelled`. After a cancelled OpenOCD run, reset the target so the board is not left halted.
+
+**Web-client:** add a "Cancel" button in `FlashProgress` while the task is pending or running.
+The `cancelled` state is already handled ("Flash cancelled", "Try again").
+
+**Done when:** cancelling a flash mid-compile stops it on the hub within a few seconds and the
+board keeps running its old firmware.
+
+### STM32 notes for whoever does items 1 to 5
+
+- The first STM32 compile on a Pi takes several minutes (the core is large). The DISC1 entry
+  allows `compile_timeout: 900`; keep `DEFAULT_TIMEOUT_MS.flash` above the largest hub timeout.
+- Images go to the cloud as base64 JSON. nginx allows `client_max_body_size 25m`
+  (`deploy/nginx/hyperloop-gui.conf`), well above an STM32F4's 1 MB of flash even with debug
+  `.elf` files.
+- `.bin` files are written at `STM32_FLASH_BASE` (`src/bench/flashing/openocd.py`). A board with a
+  different flash base needs that value per board in `boards.yaml`.
+- Boards without an ST-LINK (bare chips through the UART or USB DFU bootloader) need a new
+  flasher (`stm32flash` or `dfu-util`): a `flasher` value, a class next to `openocd.py`, and
+  registry entries. Item 3 makes that a hub-only change.
+
+## Other backend changes the new UI is ready for
+
+### 6. Latest health in REST
+
+**Why:** health arrives only as `health` WebSocket broadcasts, every 30 s. After a page load the
+Hubs table and hub sheet show "—" until the next one.
+
+**Change:** the cloud keeps the latest health per hub and includes it as an optional `health`
+on `HubInfo` (or sends the latest health for every hub to a browser when it connects). In
+web-client, seed `hubStore.health` from `fetchHubs`.
+
+### 7. Port and connection counts in the hub list
+
+**Why:** the Hubs page shows device counts, which today requires `GET /ports` and
+`GET /connections` for every online hub every 30 s (`deviceStore.refresh`).
+
+**Change:** add optional `portCount` and `connectionCount` to `HubInfo`, or
+`GET /api/hubs?include=ports,connections`. web-client then fetches full port lists only on the
+Devices and Flash pages.
+
+### 8. Close idle connections on the hub
+
+**Why:** "close a connection whose byte counters have not moved for 60 s" runs in the browser
+(`deviceStore.refresh({ enforceInactivity: true })`), only while someone has the Devices page
+open, and a view-only session gets 403 when it tries. The behavior was kept as it was before
+the redesign, not endorsed.
+
+**Change:** if the behavior is wanted, the hub's serial manager closes idle connections itself
+(configurable per profile, off for `pod` mode) and reports a `device_event`. Then delete
+`enforceInactivity`, `idleRemaining` and the "Quiet · closes in" hint in web-client.
+
+### 9. Terminal and chart history after a reload
+
+**Why:** the serial terminal keeps 500 lines per device and charts keep one hour, all in browser
+memory. A reload or a second laptop starts empty, and the Custom time range can only cover what
+this browser saw.
+
+**Change:** `GET /api/hubs/{hub_id}/telemetry` already returns recent telemetry the cloud
+keeps. Add `portId` and `since` filters. In web-client, when a device is subscribed, fetch its
+recent telemetry and feed it through `telemetryStore.processTelemetry` before live data.
+
+### 10. Custom chart schemas used for parsing
+
+**Why:** schemas created under Telemetry → Schemas are saved to `localStorage`
+(`customChartSchemas`), but `detectSensorType` in `src/services/sensorParser.ts` only reads
+`src/config/sensor-mappings.json`, so a saved schema never produces a chart. This predates the
+redesign.
+
+**Change (web-client):** include `getCustomSchemas()` in detection, custom schemas first
+(`getAllSchemas` in `src/lib/customSchemas.ts`). Later, store schemas in the cloud so the team
+shares them instead of each browser having its own.
+
+### 11. Self-host Uber Move
+
+The font stack is `'Uber Move', 'Uber Move Text', 'UberMove', Archivo, system-ui, …`
+(`--font-sans` in `src/index.css`). Uber Move is proprietary and on no public CDN, so today it is
+used only on machines that have it installed; everyone else sees Archivo (loaded from Google
+Fonts in `index.html`), which the design was drawn with.
+
+To serve it to everyone, with a license that allows web embedding:
+
+1. Put the `.woff2` files in `web-client/public/fonts/`. Do not commit them unless the license
+   allows it in this repository; otherwise have CI copy them in before `npm run build`.
+2. Add to `src/index.css`, above `@theme`, one block per weight used (400, 500, 600, 700, 800):
+
+   ```css
+   @font-face {
+     font-family: 'Uber Move';
+     src: url('/fonts/UberMove-Bold.woff2') format('woff2');
+     font-weight: 700;
+     font-display: swap;
+   }
+   ```
+
+3. Remove the Archivo request from `index.html` once nothing falls back to it.
+
+### 12. Frontend polish (no API change)
+
+- **Code splitting.** The main chunk is about 1.6 MB (gzip 510 kB), mostly CodeMirror, jsPDF and
+  html2canvas. Load the Flash editor and the PNG/PDF export with dynamic `import()`.
+- **Phones.** The Hubs and Devices tables scroll sideways, as in the mockup. A stacked layout
+  below 640 px would read better on a phone at the track.
+- **Tests.** There are none beyond type-check and lint. Add Vitest for `src/lib/format.ts`, the
+  chart geometry in `ChartPanel.tsx`, and the stores.
+- **Mock.** Send a `health` message when a mock socket connects so `dev:mock` shows meters at
+  once, and add `stage`/`progress` once item 2 lands.
+
+## Checking a change
+
+```bash
+# web-client
+npm run type-check && npm run lint && npm run build
+npm run dev:mock        # in-browser fake hubs; log in with any NetID and "hyperloop-dev"
+
+# end to end, against real code paths: cloud-services + rpi-hub-server with HUB_PROFILE=dev-sim
+```
+
+For STM32 changes, finish on real hardware: a DISC1 on a lab hub, flashed once with `.ino`
+(compile path) and once with `.bin` (OpenOCD only), then "Open terminal" to see the new firmware
+print. See `.claude/rpi-hub-setup.md` for the hub side (OpenOCD install, `lsusb` checks).
diff --git a/cloud-services b/cloud-services
index 46e47f2..a10436a 160000
--- a/cloud-services
+++ b/cloud-services
@@ -1 +1 @@
-Subproject commit 46e47f2b9fc7f7179bef359bf4480d8914a84823
+Subproject commit a10436aff2516c0e1497b21743fabe0d2387d93e
diff --git a/web-client b/web-client
index 1cd2f53..4da6c21 160000
--- a/web-client
+++ b/web-client
@@ -1 +1 @@
-Subproject commit 1cd2f53483365ac3755ffa75d419c27fa89d7cd6
+Subproject commit 4da6c21c0f934093b06b0bdfbb1748304f27eb98