[quickbeard]

Stacked on #227.

Problem

Every backend parked the guest-loopback broker stub on port 8788. That's fine for vz/bwrap — their guests have private loopbacks — but all WSL2 distributions share one VM loopback: with two Windows sandboxes running, the second socat lost the 8788 bind silently (backgrounded with &), and the second sandbox's model traffic landed on the first sandbox's broker, where the per-start auth secret rejected it with 401s.

Fix

• The WSL provider reserves a per-sandbox stub port with the same reserveLoopbackPort() used for the server port — the Windows loopback is the right uniqueness domain because WSL republishes guest loopback binds there.
• resolveEgress now receives { guestBrokerPort } so the desktop's gateway guest config points the model provider at the actual stub. vz passes its image-baked 8788 (codev-init), bwrap its GUEST_BROKER_PORT — no behavior change on those backends.
• e2e-wsl.ts gains --name for concurrent instances; each instance's broker upstream returns a per-name marker, so cross-sandbox leakage shows up as a wrong marker instead of a false pass, and the guest probe discovers the stub port from the running socat instead of assuming 8788.

Verification

Unit: bun test src/wsl/ (egress test now asserts the port handed to resolveEgress is the one socat listens on, and that it isn't the shared 8788); typecheck clean in codev-sandbox and desktop.

Live, two concurrent sandboxes on Windows 11 / WSL 2.7.11 (--name alpha --hold + --name beta):

alpha: broker via cgroup: OK   (marker broker-alpha)
beta:  broker via cgroup: OK   (marker broker-beta)   ← ran while alpha held
beta:  firewall: OK, e2e OK
alpha re-probe after beta's full lifecycle:
  port:51582
  BROKER:{"marker":"broker-alpha"}

Before this fix, beta's stub would have died at bind time and its probe would have hit alpha's broker.

🤖 Generated with Claude Code