Fix MNR noise tracking and redundant activation control

[rfoust]

Summary

Fixes #4806.

MNR's minimum-statistics estimator took a raw minimum across 25 periodograms, which systematically underestimated stationary noise power and left the Wiener gain near unity. Its decision-directed recurrence also mixed raw FFT power into a dimensionless SNR term, and its strength-blended output gain fed back into the adaptive state. Together those defects limited maximum-strength stationary-noise attenuation to about 4.54 dB.

This change smooths the periodogram before minimum tracking, applies a calibrated minimum-statistics bias, gives upward noise updates a controlled rise, stores the previous posterior SNR in the decision-directed recurrence, and keeps the adaptive gain independent of the user strength blend. The first real-energy frame seeds the estimator so the synthetic latency-prefill frame cannot poison startup or reset.

The redundant Enable MNR checkbox and its dead MnrEnabled settings/signal wiring are removed. The MNR method selector is now the sole enable/disable control, consistent with the other AetherDSP methods. The strength slider remains and now documents 0 as bypass.

Constitution principle honored

Principle XI — Fixes Are Demonstrated. The correction includes deterministic quantitative DSP coverage plus authenticated, RX-only agent automation bridge proof of the production UI and audio path.

Root cause and correction

• Smooth each FFT-bin periodogram before the 25-frame minimum so random near-zero raw samples do not masquerade as the noise floor.
• Bias-correct the smoothed minimum and limit upward tracking so the estimator adapts without immediately chasing wanted speech.
• Feed the previous dimensionless posterior SNR—not raw FFT power—into the decision-directed a-priori SNR calculation.
• Maintain the full adaptive Wiener mask internally and apply the strength control only during synthesis.
• Ignore the all-zero latency-prefill analysis frame and seed all history slots from the first frame containing real energy.
• Continue applying one shared mask to independent left/right synthesis, preserving stereo balance.
• Remove the legacy MNR checkbox, dead flat settings write, and pass-through activation signal.

Test plan

• Local build passes (/opt/homebrew/bin/cmake --build build-codex-arm64 -j8; one Apple linker process crashed under parallel test linking, and the incremental serial retry completed successfully)
• Behavior verified on a real radio if applicable (RX-only FLEX-8600 session; reporter tested the final build and approved publication)
• Existing tests pass (CI)
• Reproduction steps documented if user-reported bug

Local verification:

• Native ARM64 configure and complete build; host/system processors and final Mach-O executable are ARM64, with no RNNoise x86 source build statements.
• ctest -R 'mac_nr_filter_test|mono_dsp_stereo_adapter_test|aether_dsp_mode_policy_test': 3/3 passed.
• Stationary white-noise attenuation: -24.42 dB at input RMS 0.025, 0.050, and 0.100 (scale-invariant within 1 dB), versus the reproduced -4.54 dB baseline.
• Speech-like multitone plus noise: input SNR 1.93 dB, output SNR 20.69 dB, +18.77 dB improvement, with -2.82 dB desired-signal level change.
• Strength 0: correlation 1.000000 and 0.000 dB level change after fixed latency.
• Shared-mask stereo check: exact 4.0000 L/R RMS ratio retained.
• Startup silence remains finite/zero; post-reset stationary-noise attenuation is -24.48 dB.
• tools/check_engine_boundary.py --strict: 0 blockers (known baseline warnings only).
• tools/check_a11y.py: no new findings (existing repository warnings only).
• git diff --check: clean.

Agent automation bridge proof

The final ARM64 app was launched with authenticated automation in an exclusive radio-lock session; no TX controls were invoked.

1. Initial get dsp: active method NR2; MNR.available=true, MNR.enabled=false.
2. dumpTree: visible dspMethodBtnMNR and mnrStrengthSlider; no Enable MNR checkbox. Slider tooltip reports 0 = bypass, 100 = maximum.
3. One invoke dspMethodBtnMNR click, followed by get dsp: active=MNR, MNR.enabled=true, and NR2 disabled.
4. A second click, followed by get dsp: active=none, MNR.enabled=false.
5. Strict audioCapture probeDspStereo MNR: ok=true, audible=true, fullCoverage=true, exact 4:1 stereo ratio preserved, and both channels measured the same 0.0587033 output/input level ratio.

Proof

MNR selected at maximum strength with the redundant checkbox removed:

Checklist

• Commits are signed (docs/COMMIT-SIGNING.md)
• No new flat-key AppSettings calls — the dead MnrEnabled write is removed
• Code is clean-room — not decompiled, disassembled, or reverse-engineered from a proprietary binary (Principle IV)
• All meter UI uses MeterSmoother (no meter UI changed)
• User-visible behavior is documented in this PR and the in-app MNR description; CHANGELOG.md is untouched
• Security-sensitive changes reference a GHSA if applicable (not applicable)

Generated with OpenAI Codex.

[Copilot]

Pull request overview

This pull request addresses weak stationary-noise suppression in the macOS minimum-statistics noise reducer (MNR) by correcting the estimator and decision-directed SNR recurrence, and simplifies the UI by removing the redundant “Enable MNR” checkbox so the DSP method selector is the sole enable/disable control.

Changes:

• Fixes MNR noise-floor tracking and decision-directed update logic (smoothed periodograms, calibrated min-statistics bias, controlled upward noise rise, correct posterior-SNR recurrence, and strength blend applied only at synthesis).
• Removes the legacy “Enable MNR” checkbox and its dead settings/signal wiring; retains the strength slider with clarified “0 = bypass”.
• Adds a deterministic macOS-only regression/quality test target and wires it into CTest.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/mac_nr_filter_test.cpp	Adds deterministic offline DSP coverage for MNR attenuation, scale invariance, bypass correctness, stereo balance, startup/prefill, and reset behavior.
src/gui/MainWindow_Wiring.cpp	Removes redundant wiring for the deleted MNR enable checkbox signal.
src/gui/AetherDspWidget.h	Removes the MNR enable checkbox member and its signal declaration.
src/gui/AetherDspWidget.cpp	Removes the checkbox UI/settings write; improves strength slider tooltip and a11y metadata; keeps strength synchronization.
src/gui/AetherDspDialog.h	Removes the dialog-level relay signal for the deleted MNR enable checkbox path.
src/gui/AetherDspDialog.cpp	Removes the connection relaying the deleted widget signal.
src/core/MacNRFilter.h	Updates the filter interface/state for corrected estimator behavior and synthesis-time strength blending.
src/core/MacNRFilter.cpp	Implements smoothed minimum-statistics tracking, corrected decision-directed recurrence, prefill-safe initialization, and strength-independent adaptive state.
CMakeLists.txt	Adds mac_nr_filter_test (Apple-only) executable and registers it with CTest.

[aethersdr-agent]

Solid work, and the diagnosis is right on all three counts — the raw 25-frame minimum really was under-reporting stationary noise, m_prevPow really was feeding raw FFT power into a dimensionless recurrence, and folding the strength blend back into m_prevGain really did contaminate the adaptive state. Storing the previous posterior SNR and keeping m_filterGain strength-free are both textbook-correct now, and moving the blend into synthesizeFrameWithCurrentGain() is the right seam. I traced the UI removal through the tree: nothing reads the flat MnrEnabled key (persistence goes through ClientMnrEnabled), the method selector at AetherDspWidget.cpp:378 already calls setMnrEnabled(), the non-macOS gate at line 239 still carries the "macOS only" tooltip, and the new resetCurrentTab behaviour now matches every other tab. Clean removal. The new deterministic test target is a genuinely good addition — CMake placement matches its if(APPLE) siblings, and all four checks are green.

My one substantive question is about tuning rather than structure: OVER and FLOOR were almost certainly chosen to compensate for the estimator that this PR just fixed, and the numbers in your own test plan suggest they're now double-correcting.

Would like addressed before merge

• MacNRFilter.h:71 — OVER = 2.0 on top of a now-calibrated noise floor pins the mask near FLOOR for any stationary noise; worth a deliberate answer rather than leaving it implicit.

Polish

• tests/mac_nr_filter_test.cpp:199 — the desired-signal-loss assertion has 0.18 dB of headroom against your own measurement; likely to flake on a different Accelerate build.

Non-blocking notes

• src/core/MacNRFilter.cpp:144 — the zero-frame poisoning you fixed at startup can still recur mid-stream after a silence gap; the m_noiseInitialized seed only covers frame zero.
• The class doc block at the top of MacNRFilter.h (outside the diff) still describes the old design — "25-frame noise history … vs 10 frames before", no mention of the smoothed periodogram, bias calibration, or rise limiting. Worth a refresh while you're in there, since that comment is the first thing the next reader hits.

---

_{🤖 aethersdr-agent · cost: $8.2592 · model: claude-opus-5}

[aethersdr-agent]

Nice piece of DSP work — the core diagnosis holds up on inspection. The decision-directed recurrence is now the textbook form (ξ = α·G²·γ_prev + (1-α)·max(γ-1,0)), where the old code was feeding raw FFT power in for γ_prev, which made the estimator scale-dependent; and pulling the strength blend out of m_filterGain so it only lands in synthesis is the right separation. I also confirmed the claim about the dead setting: nothing reads the flat MnrEnabled key — the persisted state is ClientMnrEnabled, written by MainWindow.cpp:3630 and read at MainWindow_Wiring.cpp:1677 — and the method selector (AetherDspWidget.cpp:378) already owns enable/disable, so the checkbox really was redundant. No stale Enable MNR references remain anywhere in the tree, and the new <algorithm> include in the header fixes a genuine transitive-include dependency for std::clamp in setStrength().

Nothing here blocks merge. Two estimator-seeding edge cases and a CI-coverage question are worth a look.

Polish

• Seeding m_noiseEst at the full power of the first energetic frame means enabling MNR (or a reset() on band change) mid-signal briefly suppresses wanted audio rather than passing it through — inline.
• The > 0.0f energy test accepts a single denormal, so a near-silent-but-dithered first frame can seed a ~zero floor that takes the history flush plus the NOISE_RISE ramp to recover from — inline, with a suggestion.

Non-blocking notes

• mac_nr_filter_test is registered only under if(APPLE) — inline; worth confirming the macOS CI job actually builds it, given the local linker crash noted in the test plan.
• m_frameCount (MacNRFilter.cpp:275, .h:97) is incremented and reset but never read — pre-existing dead state, and this PR is already tidying the estimator fields, so it'd be a natural thing to drop while you're in here. Outside the diff, hence not inline.

---

_{🤖 aethersdr-agent · cost: $4.9903 · model: claude-opus-5}

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated no new comments.

Suppressed comments (1)

tests/mac_nr_filter_test.cpp:299

• In the reset/noise-history test loop, processed is reinterpreted as float* and appended without validating that MacNRFilter::process() returned the expected byte count. If process() ever regresses and returns a shorter/longer buffer, this will read past the end of processed and can crash the test instead of failing cleanly.

    for (int frame = 0; frame < kResetFrames; frame += kBlockFrames) {
        QByteArray block(kBlockFrames * 2 * static_cast<int>(sizeof(float)),
                         Qt::Uninitialized);
        std::copy_n(input.data() + 2 * frame, kBlockFrames * 2,
                    reinterpret_cast<float*>(block.data()));
        const QByteArray processed = filter.process(block);
        const auto* values = reinterpret_cast<const float*>(processed.constData());
        output.insert(output.end(), values,
                      values + processed.size() / static_cast<int>(sizeof(float)));
    }

[jensenpat]

Reviewed final state at 590c06e: all prior review points addressed in code (documented OVER rationale, -4 dB loss guard, MIN_FRAME_POWER silence freeze, 0.25 conservative seed, macOS CI runs mac_nr_filter_test, dead m_frameCount dropped). DSP recurrence and strength/mask separation verified; no stale MNR checkbox references remain. All checks green.