Add recorder: raw CAN and GNSS sessions with failsafes and a storage bound

Reads SocketCAN and gpsd directly. Sessions start on the first frame or on
'ctl start', end after 15 s of bus idle or on 'ctl stop'; blips under 3 s are
discarded. candump -L chunks fdatasync'd every second and zstd'd once closed;
oldest sessions deleted to stay under 50 GB / above 20 GB free. Kernel and
queue drops, link changes and clock steps are recorded as events.

Bench-tested on vcan with the first car capture (byte-identical) and installed
on the car Orin as yukart-recorder.service.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DDwfd6Ki8UfG4hK6W5YpJz

docs/impl/plans/INITIAL.md

@@ -0,0 +1,108 @@
+---
+type: plan
+status: active
+updated: 2026-10-09
+---
+# Initial plan
+
+## Order
+
+1. **`recorder`** — raw CAN and GNSS to disk. Built first because every later
+   piece is developed against what it records.
+2. `obd-poller` — standard OBD-II PIDs at a low rate, the cross-check for
+   reverse-engineered signals. The first transmitter: `can_pcan` leaves
+   listen-only for it, a deliberate step.
+3. `can-bridge` and `gnss` — publish onto the broker once it exists (project-yukart
+   Phase 3; MQTT at ~890 frames/s is still unproven).
+4. `decoder` — frames to named signals via the DBC in `contracts`.
+
+## Recorder
+
+Reads SocketCAN and gpsd directly. It is tier 1, and the broker is a single point
+of failure that raw capture should not share. When `can-bridge` exists, the
+recorder still owns raw capture; it subscribes to the bus only for decoded
+signals.
+
+### Session triggers
+
+| Trigger | Starts on | Ends on |
+|---|---|---|
+| auto | the first CAN frame | `--idle` (15 s) with no frames |
+| manual | `ctl start [label]` | `ctl stop`, or shutdown |
+
+- **No traffic, no session.** The PEAK attached at a desk, attached to a parked
+  car with the bus asleep, or unplugged, produces nothing — the socket is retried
+  every second in silence.
+- **Cranking stays inside one session.** A bus dropout or adapter re-enumeration
+  shorter than the idle timeout is recorded as `can_link` events, not as a split.
+- **Blips are discarded.** An auto session with under `--min-secs` (3 s) of
+  traffic and no marks is deleted at close — door unlocks and brief plug-ins.
+- `ctl start` during an auto session promotes it to manual. After `ctl stop`, auto
+  start waits for the bus to go idle, so a stop sticks.
+- `ctl auto off` (or `--no-auto`) leaves only manual sessions.
+
+The power service planned in orin-nano-env-setup (`docs/impl/plans/POWER.md`)
+is the intended second client: `ctl mark crank` for a recovered input loss, and
+`systemctl stop yukart-recorder` before poweroff.
+
+### Session layout
+
+```
+/var/lib/yukart/recordings/
+  .last-seq                    high-water mark; numbers are never reused
+  000042-20261009T163736/      sequence number orders; the local time is for humans
+    session.json               status open → closed | truncated; trigger, label, host,
+                               boot_id, start/end realtime + monotonic + clock sync,
+                               frame and drop counts, end reason
+    can-0001.log.zst           candump -L, closed chunks (5 min or 64 MB), zstd
+    can-0002.log               the open chunk, plain
+    gnss.ndjson                {"rx": <local receive time>, "msg": <gpsd report>}
+    events.ndjson              marks, can_link, gnss_link, clock_step, drops, storage_bound
+```
+
+`zstdcat can-*.log.zst can-*.log | canplayer` replays a session.
+
+### Failsafes
+
+| Failure | Behaviour |
+|---|---|
+| Power loss, crash, SIGKILL | open chunk is fdatasync'd every second; at the next start the session is marked `truncated` and its chunks compressed |
+| Crash mid-compression | compression writes `.zst.tmp` then renames; the plain chunk is removed only after |
+| Adapter unplugged, link down | reopen every second; recorded as `can_link` events |
+| gpsd down | reconnect every 3 s; CAN capture unaffected |
+| Writer slower than the bus | 100 000-frame queue (~110 s); overflow counted as `queue_drops` |
+| Kernel socket overflow | `SO_RXQ_OVFL` counted as `kernel_drops` |
+| Disk filling | oldest closed sessions deleted to stay under `--cap-gb` (50) and above `--min-free-gb` (20); the open session is never deleted |
+| Control socket fails | logged; auto capture continues |
+
+Measured on the first capture: ~33 KB/s of candump text, ~5.5 KB/s after zstd —
+**about 20 MB per hour of driving**. The 50 GB default holds years of it; the cap
+is there for the failure where something else fills the disk.
+
+### Time
+
+The car unit has no RTC battery yet and can start with a stale clock that chrony
+steps later. CAN frames carry the kernel's receive timestamp (`SO_TIMESTAMP`,
+CLOCK_REALTIME), and GNSS reports carry their local receive time — **the same
+clock**, so CAN and GNSS line up within a session however wrong that clock is.
+
+To get absolute time back afterwards:
+
+- `session.json` records the kernel's sync state at start and end.
+- `clock_step` events record every step over 50 ms with its size.
+- Each TPV holds GNSS UTC beside its local receive time; their difference is the
+  clock error, plus the SIM7600's NMEA latency (no PPS: expect tens to hundreds
+  of milliseconds, constant enough to calibrate out).
+
+At 1 Hz GNSS against ~890 frames/s, alignment to a few hundred milliseconds is
+what this gives. It is enough to match CAN speed against GNSS speed; it is not
+enough to time individual frames against position.
+
+## Open
+
+- **Session marking from the app.** `app-flutter`'s mark button reaches the
+  recorder through `svc-gw`, which will forward to the control socket.
+- **Native builds.** `scripts/deploy-recorder.sh` cross-compiles from the
+  workstation while the development Orin is offline; `check-glibc.sh` gates it.
+- **Copy-off.** Recordings stay on the car until the parked-at-home step in the
+  power plan, or a manual `rsync`.