Permanent WSL2 dev fix for multi-subnet source-address trap

Mirrored mode re-clones the Windows NIC's addresses each boot, so the kernel
keeps picking the wrong source for stacked device subnets (10.0.10.x sourced
from 192.168.1.123) — ARP resolves but ping/TCP dies, and every runtime
ip-route fix is wiped by wsl --shutdown.

deploy/wsl-fix-route-source.sh pins each scope-link route's src to this
host's own address in that subnet (no hardcoded IPs, idempotent, preserves
metric, non-fatal per route, waits for the route at boot). deploy/parking-net
.service reapplies it on every boot.

Dev-box only; the appliance is bare-metal Linux with static networkd config.
Verified: camera pings with no -I flag; driver pulls a snapshot with no
localAddress set.

wiki/concepts/wsl-dev-networking.md

@@ -56,6 +56,59 @@ Mirrored networking is necessary but **not sufficient** — these still bit us:
   (`127.0.0.1`). Node's Vite proxy can stall on the v6 attempt before falling back — point the
   proxy at `127.0.0.1` explicitly. (See [[local-dev-workflow]].)
 
+## Multi-subnet source-address trap (the "ARP works but ping/TCP dies" bug)
+
+Field devices arrive **statically configured on assorted `/24`s** by whoever installed them last
+(e.g. a camera on `10.0.10.121`, a printer on `10.0.10.6`, others on `192.168.1.x`). The host
+copes by carrying **one IP per device subnet on a single NIC** (this is correct — you do **not**
+need a NIC per subnet). But stacking subnets on one interface exposes a Linux source-selection
+trap:
+
+- Connected routes come up as `proto kernel scope link` **with no preferred source**. With two
+  such subnets on one NIC, the kernel may pick the **wrong source address** — e.g. sourcing
+  traffic to `10.0.10.121` from `192.168.1.123`.
+- Symptom is baffling: **ARP resolves and the neighbor shows `REACHABLE`** (L2 is fine, source
+  address is irrelevant to ARP) while **every ping and TCP connect times out** (replies have a
+  wrong/unroutable source → dropped, possibly by uRPF). Looks like "the device is down / the whole
+  subnet is unreachable" when nothing is actually broken.
+- **Diagnose:** `ip route get <device-ip>` shows the chosen `src` — if it's an address on a
+  *different* subnet, that's the bug. Confirm by forcing the right source:
+  `ping -I <correct-src> <device-ip>` (or `curl --interface <correct-src> …`) — instant replies.
+- **Fix (runtime):** pin the preferred source on the connected route, per subnet:
+  `sudo ip route replace <subnet>/24 dev <nic> proto kernel scope link src <correct-host-ip> metric <m>`
+  (use `replace`, not `change` — `change` errors `RTNETLINK: No such file` if the route isn't up
+  yet). Do **not** delete the other subnet's address unless it's genuinely unwanted — you need all
+  of them to reach all the devices.
+
+- **Fix (permanent, this box):** `deploy/wsl-fix-route-source.sh` + `deploy/parking-net.service`.
+  The script walks each `proto kernel scope link` route on the NIC and pins `src` to THIS host's own
+  address in that same subnet — **no hardcoded IPs**, so it also covers future device subnets; it's
+  idempotent, preserves the route metric, and tolerates a missing route. The systemd unit (oneshot,
+  `enabled`) reapplies it on every WSL boot — which is the point, since `wsl --shutdown` otherwise
+  wipes the runtime fix (mirrored mode re-clones the Windows addresses fresh each boot, see below).
+  Install once: copy the unit to `/etc/systemd/system/`, `systemctl enable --now parking-net`.
+  Gotchas hit while building it: `network.target` is too early for mirrored-mode addresses (the
+  script waits up to 15s for a route to appear); and it must NOT `set -e` or one failed `ip` call
+  aborts the whole boot fixer.
+
+> **Root cause is on the Windows side.** Mirrored mode clones the Windows host NIC's addresses into
+> Linux at every boot, so the stray `192.168.1.x` lives on Windows — the truly permanent fix is to
+> remove/reconfigure it there (or set `SkipAsSource`/interface metric). The systemd hook is the
+> self-contained Linux-side answer that needs no Windows changes.
+
+Verified on hardware (2026-06-15): after the hook, `10.0.10.121` pings and the real [[lpr-camera]]
+Hikvision driver pulls a snapshot with **no** source-forcing (`localAddress` becomes optional).
+
+## On the real appliance: multi-subnet is a deployment config, not a WSL hack
+
+Production is a **dedicated hardened Linux appliance** ([[disk-os-hardening]]), so the WSL story
+above is dev-only. The device-subnet problem persists, though, and is solved the same way at the
+OS level: the appliance NIC carries **one address per device subnet**, each connected route with a
+pinned `src`, made persistent (systemd-networkd / netplan). Per the threat model this still rides
+on **[[network-isolation]]** — device subnets are isolated segments reachable only by the host.
+The long-term clean answer is to **re-IP the devices onto one planned parking-system subnet** at
+install so the host needs only one address; the multi-subnet config is what you run until then.
+
 ## Alternative if you can't use mirrored mode
 
 Windows 10 / old WSL can't do mirrored mode. Options: run the **backend natively on Windows**