[Nick Hehr (HipsterBrown)]

Why

internal/controller discarded a valid servo reading whenever the servo set any status flag. An overloaded servo (a gripper clamped on a part, a hot joint) was therefore unreadable, and the arm's dwell, IsMoving, JointPositions, motor-setup verification and arm bring-up all inherited that blindness.

feetech-servo v0.7.0/v0.7.1 fixed the library half: reads return the payload alongside a condition error, and feetech.ConditionStatus(err) says whether the payload is usable. This PR is the module half, extracted from #42 so the gripper grasp latch there can rebase onto it. It is wider than #42's two sites.

What changed

• feetech-servo v0.6.1 → v0.7.1. No signature changes. v0.7.1 also makes writes return nil on condition-only flags; nothing on main inferred "overloaded" from a write error (checked every SyncWrite/WriteRegister/RegWrite/SetGoals/gain/torque write), so this is strictly fewer spurious failures. One visible effect: applyServoGains now completes D→I→P on a flagged servo instead of aborting mid-sequence.
• One helper, tolerateCondition(err) (feetech.StatusError, error), is the single place the policy lives. Nine read paths in manager.go route through it right after their bus read: ServoPositionPercent, AnyServoMoving, SyncReadPositions, GetJointPositionsAndMovingForServos (the dwell's read), GetJointPositions, GetJointPositionsForServos, LoadForServos, Ping, PingServo. Only the first two expose the flags, because only they have a consumer; Ping logs them. ReadServoRegister and config.go's calibration reads are deliberately left intolerant.
• servo_position / servo_moving carry condition as text (servo status error: [overload]), absent when clear. Text because a protobuf Struct turns every number into a float64 on a remote arm; a field because a typed error does not survive the DoCommand boundary at all. No consumer on main until Latch the grasp instead of inferring it, and add servo_load #42 lands — it is the wire contract Latch the grasp instead of inferring it, and add servo_load #42 rebases onto.
• testfake.FakeTransport.SetStatus injects a status byte through the real protocol, so every path above is unit-tested without hardware via a clean-vs-flagged handle pair.
• Docs: docs/arm.md response samples; one CLAUDE.md gotcha.

Hardware behaviour changes to know about

• A clamped or hot gripper no longer fails arm bring-up (doServoInitialization → Ping) or reads as not_responding in motor_setup_verify.
• feetech.Discover now includes condition-flagged servos (new FoundServo.Status), so discoverOneMotor sees a hot motor it previously dropped. Direction is right; untested here.
• services/discovery's port probe still treats any ping error as "no servo" — it holds no handle. Follow-up.
• WaitForServosToStop on an overloaded-but-stopped servo now returns when Moving = 0 instead of erroring (tested against the real fake-bus path).

Testing

go test ./..., -race on internal/... and components/..., go vet, gofmt -s clean. Mutation checks executed: stubbing the helper fails all seven position/moving/load subtests plus three helper cases; un-wiring one site fails exactly that site's subtest.

Hardware test: clamp the gripper by hand, then (1) servo_position on the arm returns a reading with condition set, (2) restart the arm component with the gripper still clamped and confirm it comes up.

🤖 Generated with Claude Code