fix: make beetroot modes/doctor tell the truth across environments (#179, #191, #164)

[Xiddoc]

Three diagnostics-truthfulness fixes so beetroot modes / beetroot doctor stop endorsing hosts that can't actually boot the backend they report as healthy. Closes #179, #191, #164.

#179 — modes reported redroid host/auto "supported" with a stopped daemon

survey() probed Docker with shutil.which(docker_bin) (CLI presence only). A host with kernel binder ready but dockerd stopped (freshly-rebooted runner, rootless without a running user daemon, socket-less container) printed a false supported, and the subsequent up failed at docker compose up.

• Relocated the docker info probe out of builder.py into capabilities.py as a public, single-sourced docker_daemon_responsive() (builder now reuses it; no capabilities→builder import cycle — capabilities only imports hostcheck/settings/vm.qemu).
• classify_modes gains a pure docker_daemon: bool; _redroid_host returns supported only when binder is ready and the CLI is present and the daemon responds. CLI-present-but-daemon-down → needs-setup with a "start the Docker daemon" remedy.
• survey() populates the flag via the shared helper, short-circuiting the probe when the CLI is absent.

#191 — doctor reported vm.accel: pass with QEMU/artifacts absent

On the canonical binderless setup (before apt-get install qemu-system-x86 + beetroot build --vm-kernel), doctor printed vm.accel: pass while the VM physically couldn't boot — contradicting beetroot modes (needs-setup) and offering no remedy.

• VmDeviceBackend.health() now emits a vm.qemu row (shutil.which(settings.qemu_bin); the fail reason reuses the shared capabilities._QEMU_INSTALL string so doctor and modes agree) and a vm.artifacts row (resolves kernel + rootfs via the existing _resolve_artifact; the QemuLaunchError message already names the missing artifact and points at beetroot build --vm-kernel). vm.accel is no longer the sole readiness signal.

#164 — vm doctor false-failed a healthy micro-VM

The vm health path inherited the USB-oriented adb.serial row (_check_adb_serial_listed), which needs a serial to already be listed — but a TCP adb target only appears after an explicit adb connect, so doctor false-failed a fully-booted VM from a fresh adb-server lifetime (and exited non-zero, breaking scripts gating on beetroot doctor).

• health() now issues adb.connect = api._check_adb_connect(self.adb_address) (connect-then-verify, mirroring the redroid path) instead of adb.serial. The Frida/Magisk rows remain intentionally absent (network-isolated guest, plain redroid).

Tests

• test_capabilities.py: classify_modes for binder-ready × {daemon up, daemon down}; three survey() integration tests proving the daemon probe is threaded into the verdict (live→supported, dead→needs-setup+remedy, CLI-absent→probe skipped); the real-probe docker_daemon_responsive tests moved here from test_builder.py.
• test_vm_backend.py: happy path (qemu+artifacts+adb all pass, asserting the connect row targets backend.adb_address and that adb.serial is gone); qemu-absent → vm.qemu fail with the install remedy; artifacts-absent → vm.artifacts fail with the build hint; adb-absent → connect row skips.
• test_builder.py: updated to patch capabilities.docker_daemon_responsive (one patch covers every relocated call site).

Docs (cli.md, sandbox-quickstart, running-in-ci, binder-and-modes, the vm design log) updated to describe the new doctor rows truthfully. Full gate green locally: ruff, mypy --strict (src + tests), pytest 1983 passed at 100% line+branch coverage.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WmChW67jv2BaPk1a8HreHU

---

Generated by Claude Code