Pivot orchestrator to continuous-control VLA loop

[9LogM]

Summary

• Replaces the symbolic VLM-per-command path (Gemini + mission state machine + re-prompt loop) with a continuous-control VLA loop. Each tick the orchestrator hands the latest RGB frame + held instruction + telemetry to a VLA backend at orchestrator.control_hz and dispatches at most one tool call.
• Slims the tool schema to the action space the VLA actually emits: arm, disarm, set_velocity_ned, land, abort, kill. takeoff, goto_ned, return_to_launch, hold are gone — continuous velocity setpoints subsume them.
• Default VLA backend is a no-op so the stack boots and the control loop runs without a model wired in. Drop a model into python/services/orchestrator/vla.py (OpenVLA, π0, Octo, custom) to make it act.

What changed

• vlm_base.py / vlm_vla.py / vlm_gemini.py → vla_base.py + vla.py (Gemini deleted)
• OrchestratorMission, _TERMINAL_TOOLS, stuck-detection — removed; nothing terminal under continuous control
• Voice-arm latch is now an orchestrator-side flag set by the command thread; a VLA-emitted arm without a prior spoken arm is rejected
• Frame staleness gate (orchestrator.frame_max_age_s) idles the loop when RGB is missing/old instead of letting the model extrapolate
• Config: perception.vlm.* → perception.vla.{model_path,device}; orchestrator.{mission_max_steps,telemetry_history_seconds} → orchestrator.{control_hz,frame_max_age_s}
• GEMINI_API_KEY / gemini extra removed from .env.example, pyproject.toml, Dockerfile, scripts/preflight.sh
• TUI alarm row vlm_error → vla_error
• Deleted docs/phase-2.md (mission-state design doc, superseded by this pivot)
• Swept VLM/Gemini out of every comment, docstring, and README — git grep vlm gemini returns nothing

Diff stat: 23 files changed, +397 / -851 (orchestrator simplified significantly).

Tests

• 43/43 green excluding the pre-existing replay slow-joiner flake (test_phase2_replay.py — unrelated)
• Removed the mission-state-machine integration tests that no longer have a referent
• Added 9 continuous-control tests in test_orch_vla.py covering: setpoint dispatch, idle on no-tool-call, voice instruction propagation, abort short-circuit, arm-authorization latch, VLA-emitted-arm without voice rejection, stale-frame idle, bridge-offline survival, VLA exception survival

Test plan

• docker compose build succeeds; both images come up
• lexaire-python test suite green (pytest python/tests --ignore=python/tests/integration/test_phase2_replay.py)
• Stack boots end-to-end with perception.vla.model_path: null (no-op backend, control loop ticks idle)
• Wire a real VLA in vla.py and verify set_velocity_ned setpoints flow through to the bridge

🤖 Generated with Claude Code