ai-autopilot: sandboxed runner adapters (Docker / WebContainer / Flue)

[suleimansh]

Follow-up from the ai-autopilot epic (#97). The runner seam and the first real adapter (LocalRunner, #106) shipped. What's left are the sandboxed adapters that isolate untrusted agent code:

• Docker — full-fidelity, background/long-running; needs a Docker daemon. Shipped (DockerRunner, PR feat(ai-autopilot): DockerRunner — first sandboxed runner adapter #142; sandboxed boot-and-serve E2E, PR test(ai-autopilot): sandboxed boot-and-serve — serveCheck × DockerRunner E2E #143).
• WebContainer — instant in-browser Vike preview; needs a browser runtime. Shipped (WebContainerRunner, PR feat(ai-autopilot): WebContainerRunner (in-browser sandboxed runner) #223; headless-Chromium boot-and-serve harness under harness/webcontainer/, 15/15).
• Flue — mirror Flue's sandbox contract (in-memory / edge / container); needs a live Flue env.

LocalRunner is the reference each mirrors; FakeRunner covers tests. Docker is done; WebContainer and Flue remain infra-gated (can't be built and honestly verified without provisioning that infra first).

[suleimansh]

Docker adapter done: #142. DockerRunner boots each workspace as a container via the docker CLI, same seam as LocalRunner, verified against a live daemon (15 container tests; skips cleanly with no daemon so CI stays green).

WebContainer and Flue are still infra-gated — leaving this open to track them.