fix(controller): propagate close-state to all shared-controller holders

[Nick Hehr (HipsterBrown)]

Summary

The arm and gripper on one SO-101 share a serial port via the ref-counted controller registry. Previously each consumer got its own *SafeSoArmController struct copy wrapping the same bus. When one component closed first, ReleaseController closed the shared bus, but the other holder's struct had no way to observe it — its next call raced against / hit a closed bus.

This is the data-race fix extracted from #11 (which bundled several unrelated refactors). Only the race fix is here; the rest is intentionally left out (see below).

Changes

• manager.go — add a closed atomic.Bool to SafeSoArmController plus an ErrControllerClosed sentinel. Every public method checks it up-front and returns the sentinel instead of touching a closed bus. Close() is now idempotent (sets closed).
• registry.go — getExistingController / createNewController return the cached controller pointer instead of a fresh per-caller struct copy, so all holders share one closed flag. ReleaseController and ForceCloseController set closed=true before closing the bus, so surviving holders observe ErrControllerClosed atomically on their next call.
• Tests — a scripted mock-bus harness (mock_bus_test.go, no hardware) + regression tests (lifecycle_test.go): the gated-method sentinel across all 9 methods, same-pointer-per-port, and release-closes-all-consumers.

checkClosed is best-effort (documented in-code): callers must hold a registry refcount for the duration of a call. The runtime.Caller-based release tracking is left intact here, so the fix is self-contained.

Explicitly out of scope (deferred from #11)

• Replacing runtime.Caller release tracking with explicit port handles
• Per-call ctx threading through init/diagnose/verify
• The motion.FromProvider controller-leak fix and cancelCtx/cancelFunc cleanup
• Design-spec / implementation-plan docs

Testing

• gofmt -s clean, go vet ./cmd/module/ . clean
• go test -race ./cmd/module/ . — full so_arm package passes under the race detector

Note

The cached-pointer approach means a second component reacquiring a port with a different calibration file updates the shared calibratedServos but not the controller struct's calibration field. This is pre-existing behavior carried over from #11's design, unrelated to the race, and only affects the rare reacquire-with-different-calibration path — flagged for awareness.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JNvy6dGUT6wBHLNBv8YPYv