OBR head-tracking and limiter setters share the render mutex with Process()

[trsonic]

Tracking issue for the design proposed by @olilarkin in #27, which is closed without merging (see below).

Problem

Since #25, SetHeadRotation(), EnableHeadTracking() and EnableLimiter() in src/renderer/obr/obr_capi/obr/obr/renderer/obr_impl.cc take mutex_, the same absl::Mutex that Process() holds for the whole render call. #25 fixed a real race: a four-float quaternion written concurrently with Process() could be read torn. But the fix puts the control thread and the audio thread on the same lock.

absl::Mutex has no priority inheritance. A control thread preempted while holding mutex_ stalls rendering until the scheduler runs it again. Head rotation is written continuously from a sensor thread, so the lock is contended on every buffer, not occasionally. The audio thread should not wait on a lower-priority thread.

Proposed fix (by @olilarkin, #27)

Publish the three values instead of locking them:

• EnableHeadTracking() and EnableLimiter() flags become std::atomic<bool>.
• The quaternion goes through a seqlock over four std::atomic<float> (AtomicWorldRotation in obr/common/atomic_rotation.h). std::atomic<WorldRotation> is 16 bytes and not lock-free on the platforms obr targets, so it would reintroduce a hidden lock on the audio thread.
• The seqlock read is bounded, not spin-to-convergence. After a few failed attempts Load() returns the caller's previous value, so the read is wait-free. Reusing one buffer's rotation is inaudible; missing a deadline is not.
• Process() latches all three values once at the top of the block. ProcessingGroup interpolates the rotation across the buffer, so a value that changes partway through is already wrong.
• mutex_ still guards audio_elements_ and processing_groups_, which are rebuilt rather than handed over. Nothing that runs continuously during playback takes it.

Tests in the proposal: atomic_rotation_test (round trip, no tearing under concurrent writes, bounded fallback under a write storm) and ObrImplTest.TestConcurrentHeadRotationDuringProcessing (render while a second thread sweeps rotation and toggles both flags). Both clean under ThreadSanitizer; the existing renderer and processing-group tests pass unchanged. No public API change.

The implementation is on SpatialAudioKit/oar:lock-free-control-state (33c4b09, 8 files, +379/-24).

Status

#27 is closed without merging because the Contributor License Agreement was not signed. @olilarkin agreed in the PR thread that the contribution can be adapted. The fix will be resubmitted as a fresh PR based on this design, rebased onto current main, with credit to @olilarkin for the analysis and the design.

Related: #28, the per-render malloc/free in Process(), noticed while measuring this change and filed separately.