Sync post-M145 upstream fixes (full parity with C++ reference)

[dignifiedquire]

Summary

This PR syncs sonora with the post-M145 upstream WebRTC fixes, for full parity between the Rust port and its C++ reference. The Rust port gets the upstream changes it was missing. The cpp submodule moves to a fork branch with the same commits cherry-picked, so the Rust-vs-C++ comparisons stay in lockstep.

Important

Depends on dignifiedquire/webrtc-audio-processing#1. The submodule points at fork commit 3b33b6e. Merge the fork PR with a merge commit first; squash or rebase would orphan 3b33b6e.

User-visible change: IVC defaults

InputVolumeControllerConfig::default() held test-style values that matched neither M145 nor upstream. It is now upstream's production Config{} after d9b92fe1bc:

Field	Before	After
Target range (dBFS)	[−30, −18]	[−50, −12]
enable_clipping_predictor	false	true
update_input_volume_wait_frames	0	100
Speech probability / ratio thresholds	0.5 / 0.8	0.7 / 0.6

With input_volume_controller enabled, this changes recommended_stream_analog_level(). Whether the processed samples change depends on what happens to that recommendation:

• Fixed analog level (the caller ignores the recommendation): samples are unchanged.
• The caller applies the recommendation to its microphone or via set_stream_analog_level(), or capture_level_adjustment.analog_mic_gain_emulation is enabled (the emulated analog gain follows the recommendation): the output level changes. A probe found 1.1–1.2 million of 1.44 million samples changed.

There's a CHANGELOG entry under "Unreleased".

Upstream commits

• Ported to Rust (9):
  • a6d83994aa: stale AECm documentation only; the port never had AECm.
  • 7f5a8b656c
  • d460e60e19: headroom fix. The test that asserted the buggy value now asserts 0.
  • d9b92fe1bc
  • e10cd19640: the non-neural-estimator parts, including the MovingAverage start-up normalization and the config_changed flow.
  • fe9c0a1de4, 4252c6051d, 8dd6b47c83
  • 86ef7fa42d
• Already in Rust, now also in C++ (6):
  • 297352a2fd, 573e746914 and 7c388cbabb, via 9e4b402. That commit's message swaps the first two hashes.
  • 32feae39d0, 05d6bc269c and 401392b7e6: the port never had these field trials.

Commits

• bf41756 port: Remove aec_mobile (AECm) config from audio_processing.h
• 3ea7be9 port(aec3): Propagate delay estimator fields to AecState.
• 9e455f7 port(aec3): AEC3: Fix render buffer headroom calculation on underruns
• d09e7ff port(agc2): Make max target input level for input controller -12dB
• dce1a5c port(aec3): Audio: Enable dynamic on-the-fly AEC3 configuration updates for the suppressor gain computation
• b4fe2fc port(aec3): Remove the usage of the field-trial WebRTC-Aec3AecStateFullResetKillSwitch
• cb89e6d port(aec3): Remove the usage of the field-trial WebRTC-Aec3CoarseFilterResetHangoverKillSwitch
• 946592e port(aec3): Remove the usage of the field-trial WebRTC-Aec3DeactivateInitialStateResetKillSwitch
• 492e58d port(aec3): aec: fix wrong sentinel usage in echo metrics
• 8149e8c chore(cpp): update submodule with post-M145 upstream cherry-picks
• 1629a49 test: compare 48 kHz stereo echo pipeline with decorrelated channels against C++
• 1147f9c docs: note post-M145 upstream cherry-picks
• f556978 test(aec3): check that comfort noise is identical on all channels
• 4a021c1 test(aec3): port SuppressionGainTest.BasicGainComputation and cover config_changed
• c1375c1 test(agc2): build IVC tests on the production config where upstream does
• 45cad1a fix(sonora-sys): rebuild the shim when the C++ headers or library change
• 5dc45e4 docs: keep the 0.1.0 changelog entry and name sonora-ffi as the backend
• 024f9ce test: give the stereo echo comparison a tolerance that catches mono render
• 8a829c4 chore(cpp): update submodule with the AVX2 reference and NEWS fixes
• 828cc67 docs: note that the neural residual echo estimator is not ported
• 7a129c1 docs: correct the IVC and AEC3 changelog claims and the ML-REE notes
• 72366c3 docs: say how the IVC recommendation reaches the output

Tests added or strengthened

• stereo_echo_pipeline_matches_cpp (sonora-bench): 48 kHz stereo with different signals on each channel, against C++, with tolerance 0.02. It fails if render or capture falls back to mono. It cannot detect a regression of the comfort-noise or filter-switching changes, which move the output by less than 1e-5; its doc comment says so.
• Comfort noise: a test that fails if the channels draw independent random phases again.
• SuppressionGainTest.BasicGainComputation: ported, plus a config_changed test. Deliberately breaking the code in three ways each made a test fail.
• IVC tests: those that upstream builds from Config{} now use the production default.
• sonora-sys build: build.rs now reruns when the C++ headers, the installed library, or WEBRTC_CPP_ROOT change. Before, the shim could silently link against stale headers.

Verification (local, macOS arm64)

• cargo fmt --all --check, and both CI clippy -D warnings commands
• cargo test --workspace: 772 passed, 0 failed, 0 ignored
• RUSTDOCFLAGS=-D warnings cargo doc --workspace --no-deps, and cargo +1.91.1 check --workspace --all-targets
• sonora-bench against the rebuilt C++: 27 passed, 0 failed
• C++ suite with the Rust backend (CI flags and filter): 2327 passed, 0 failed. Also 0 failed under emulated x86_64.
• Real x86 hardware and GitHub CI: this PR's run covers both.

Known and deferred

• Lost upstream format test: the submodule bump takes upstream 1d1ee4503c, which turns AudioProcessingTest.Formats (53 format-conversion cases) into an error-code-only test. That removes the only Rust-vs-C++ output check across rate and channel conversion; sonora-bench covers only matched formats. Follow-up: add sonora-bench comparisons for conversions (for example 48k stereo to 16k mono, 44.1 kHz, mono to stereo).
• Input volume controller tests: the production defaults now run clipping prediction, the 100-frame wait and the 0.7/0.6 thresholds, but upstream's controller-level tests for those paths are not ported, and there is no Rust-vs-C++ comparison for the controller. This gap predates this PR. Follow-up.
• ML-REE: the neural residual echo estimator is not ported; the sonora-aec3 README and the top-level README say why.
• There is no Rust-vs-C++ comparison of the IVC default. The C++ IVC suites are compiled out under the Rust backend, and one would need new shim functions.
• There is no targeted test for the joint coarse/refined filter choice: it is inline in EchoRemover::process, as it is upstream.
• An older difference: 48k_*/ec_only differs from C++ by 0.156 against a tolerance of 0.2. The number is identical before this branch, and it hasn't been investigated.
• CHANGELOG still has the stale "0.1.0 (unreleased)" heading and no 0.2.0 entry. This predates the branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01H29DamSLugosGXJSz1e5Yx

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 96.64694% with 17 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.69%. Comparing base (413f64e) to head (72366c3).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #36      +/-   ##
==========================================
+ Coverage   93.64%   93.69%   +0.04%     
==========================================
  Files         146      146              
  Lines       31130    31484     +354     
  Branches    31130    31484     +354     
==========================================
+ Hits        29153    29499     +346     
- Misses       1753     1762       +9     
+ Partials      224      223       -1     

Flag	Coverage Δ
rust	93.69% <96.64%> (+0.04%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jacksongoode]

Hey @dignifiedquire , cool to see your work on this project!

I was curious, do you imagine you will support the neural aec backend within sonora? Or is it a bit too complicated outside the scope here?