fix(harbor): make generated tasks safe on remote sandboxes (e2b)

[vaibhavdabas16]

Advances #331 §3. Does not close it — the end-to-end -e e2b smoke run is still open work, and I have no e2b account to do it with.

Scoping

#331 says §1–§2 are largely done on feat/harbor-hub-registry and that "§3 (e2b) and the parity run in §4 are the open work", so this deliberately touches neither the task format nor the registry — no conflict with that branch. Of §3's checklist:

• ✅ No dependency on local podman/docker specifics; healthcheck-driven readiness for Chromium + CDP + interceptor
• ✅ Credentials via --ve/--env-file only; resource floor documented
• ✅ Prior harbor-0.20-port-notes.md conclusion (E2B needs no local disk) folded into docs/harbor.md
• ❌ harbor run ... -e e2b smoke task — not done, needs an account
• ❌ Prebuilt public runtime image — belongs to §2 on the hub-registry branch

The bug this fixes

Setup accepted any 200 from /api/status as "runtime ready":

if curl -sf http://127.0.0.1:7878/api/status >/dev/null \
  && curl -sf http://127.0.0.1:9223/json/version >/dev/null; then

But /api/status starts answering as soon as uvicorn binds, while eval_interceptor_ready only flips to true after the CDP handler connects and sends Target.setAutoAttach (runtime-server/server.py:296) — a loop that retries on 1s sleeps. A task that starts in that window runs with interception inactive and cannot score Stage 1, and the result reads as an agent miss rather than a harness fault. The window is small on a warm local daemon and much wider on a cold remote sandbox, which is why it surfaces here.

task.toml's [steps.healthcheck] already checked the right condition. The two now share one runtime_ready_command(), so they cannot drift. That also fixed a latent escaping trap: the healthcheck string is written into a TOML basic string where \" unescapes to ", but the same literal pasted into a shell script keeps the backslash and the grep would never match. The shell form is now canonical and task_toml escapes it once with json.dumps — the emitted task.toml is byte-identical to before (verified against main).

Also raises the setup wait from 60s to a configurable 180s, and makes a timeout tail the runtime-server log instead of printing one opaque line.

Fixed sleeps → polling

start-runtime.sh slept 1–2s after starting Xvfb, uvicorn, and Chromium. Those constants encode local daemon timing; in remote-browser mode the script declared the CDP bridge ready after a bare sleep 1. Each step now polls for the thing it needs with a bounded timeout that names the component that never came up. CLAWBENCH_RUNTIME_WAIT_TIMEOUT_S overrides the 60s per-component default.

Testing

Five new cases in tests/test_harbor_adapter.py: the healthcheck and setup script share one readiness contract in both local and kernel modes and are escaped correctly for their respective consumers; the setup wait is remote-sized and reports what failed; startup polls rather than sleeps (no bare sleep <n> left standing in for a readiness check); and a generated task carries no host paths and no baked secrets — every credential in task.toml is an env reference.

Full suite passes locally (280 passed, 10 skipped); ruff and pyright clean.

Not verified against a live Harbor run — these are static assertions on the generated dataset plus a shell-syntax check. Worth one local harbor run on the smoke dataset before merge, since the setup gate is now strictly stricter than it was.