Skip to content

Learned residual suppressor: measured ML benchmark, clean-licensed pipeline, and the pluggable-post chain - #33

Merged
tap merged 24 commits into
mainfrom
claude/ml-echo-cancellation-jjwdva
Aug 8, 2026
Merged

Learned residual suppressor: measured ML benchmark, clean-licensed pipeline, and the pluggable-post chain#33
tap merged 24 commits into
mainfrom
claude/ml-echo-cancellation-jjwdva

Conversation

@tap

@tap tap commented Aug 8, 2026

Copy link
Copy Markdown
Owner

What this changes

Adds a learned (neural) residual-echo suppressor as a selectable post-filter engine — tap::mu::nn_suppressor + aec_chain_nn, with aec_chain growing a pluggable Post template parameter — plus the tools/ml/ benchmark and clean-licensed training pipeline that measured the design into existence, the trained models (16 kHz benchmark v1, 48 kHz shipping v2), C ABI support, two executed comparison notebooks, and book/README/HANDOFF documentation. Companion PR: tap/MuTap-Max (the mutap.aec~ @postfilter 0/1/2 + @model surface).

Why

The originating question: is there an ML approach to echo cancellation that is not IP-encumbered? Measured answer: end-to-end neural AECs are (DTLN-aec deletes out-of-domain near ends at 0–5 dB SDR; its weights carry mixed-license training data), but a small learned band-gain post-filter over MuTap's own linear canceller wins single-talk ERLE at equal transparency — and every input to its training is MIT / CC BY 4.0 / synthesized here, so the weights are clean to ship. The post stage is also the one classical stage that is a heuristic, i.e. the one place a learned rule plausibly beats DSP.

Verification

  • Classical path bit-identical: tests/branchless_parity_check.cpp fingerprint d515415bee83b420 before and after the Post template refactor, both MUTAP_SUPPRESSOR_BRANCHLESS values.
  • Full gtest suite: 184/184 passed after the chain refactor; a final rerun with all subsequent (additive) commits is in flight and CI is the authoritative gate for this PR.
  • New tests: tests/test_nn_suppressor.cpp (transparency at forced unit gains to −140 dB, silence at zero gains, comfort-noise floor level, echo_explained extremes, geometry rejection, chain composition).
  • Python↔C++ parity: tools/ml/test_parity.py, 2.3e-8 relative on identical frames; re-verified with the shipped weights (1.0e-7).
  • Performance numbers: all measured, carried by the executed notebooks (notebooks/ml_aec_comparison.ipynb, notebooks/aec_head_to_head.ipynb incl. the 48 kHz shipping-geometry section) and tabulated with provenance in tools/ml/README.md. Headline (48 kHz speech, medians, C ABI chain): learned v2 53.4 dB single-talk ERLE / 34.3 dB near-end SDR vs classical 34.5 / 30.3, double-talk within 1.5 dB; off-domain music double-talk 4.0 vs 9.1 dB — classical remains the default engine.
  • Not done here: ITU battery run against the learned engine (filed as a follow-up characterization); M55/Hexagon builds of nn_suppressor are untested on target (host float32 instantiation is exercised; int8/CMSIS-NN filed).

Notes for the reviewer

  • Contract change. aec_chain gains a third template parameter with a default — source-compatible for every existing consumer (aec_chain<double> unchanged, verified by fingerprint); residual_suppressor untouched. New MUNN weights format is versioned (0002 self-describing geometry; 0001 still loads).
  • Notebooks re-executed, committed with outputs; the numbers quoted above come from their cells.
  • Licensing: DTLN-aec pretrained weights and the AEC-Challenge corpus are used benchmark-only and are not redistributed; the shipped pretrained/*.munn weights were trained exclusively on LibriSpeech (CC BY 4.0) + material synthesized in-repo. Survey in tools/ml/README.md.
  • tools/ml/ is host-side tooling behind MUTAP_BUILD_ML_TOOLS (default OFF) — nothing enters the emulated-target builds.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg

Generated by Claude Code

claude added 24 commits August 7, 2026 17:26
@claude
Add ML AEC benchmark harness: scenario dump + shared offline meter
58a5c4a 
tools/ml: a scenario dump (the test rig's double-talk protocol plus a
near-end-only transparency segment, emitted as raw float64 via the rig's
own generators and simulator) and a Python harness that runs MuTap's
shipped engines (via the C ABI) and DTLN-aec (MIT, Westhausen & Meyer
2021) on identical signals with one shared meter: delay-compensated
ERLE, double-talk true-echo suppression, and near-end preservation SDR.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add clean-licensed dataset generator + speech benchmark scenarios
18f69de 
tools/ml/make_dataset.py synthesizes echo scenarios from LibriSpeech
(CC BY 4.0): random rooms, bulk delay, speaker saturation, SER sweep,
gated double-talk — then runs MuTap's linear Kalman canceller via the
C ABI and emits (features, oracle band gains, loss weights) shards for
the learned residual suppressor. tools/ml/features.py is the single
source of truth for the analysis geometry (128/64 sqrt-Hann STFT, 22
ERB-spaced bands) that the trainer and the C++ inference must share.
A --scenarios mode emits protocol-layout speech scenarios so
run_benchmark.py measures speech material with the same meter
(--scenario-dir).

Measured on three speech scenarios (medians): DTLN-aec-512 43.4 dB
single-talk ERLE / 14.2 dB double-talk suppression / 29.3 dB near-end
SDR; mutap-chain 30.9 / 17.2 / 38.2 dB; on the rig's synthetic
materials DTLN collapses to ~0 dB suppression and 0..5 dB near-end SDR
(out of its speech-trained domain) while the classical chain holds.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add suppressor trainer, numpy reference inference, and tools/ml README
f9dc810 
train_suppressor.py: dense64-tanh -> GRU96 -> dense22-sigmoid (~51k
parameters), loss-weighted MSE on sqrt-compressed oracle gains. nn.py:
dependency-free numpy reference inference (the contract the C++ header
must match) plus the mutap-kalman+nn benchmark system, wired into
run_benchmark.py as --nn-weights. README.md: the licensing map this
effort exists to establish, the measured baseline tables, and the
reproduction recipe.

Oracle sanity bound on a speech scenario (upper bound for the learned
gains): 67.8 dB single-talk ERLE / 16.5 dB double-talk suppression /
59.3 dB near-end SDR, vs 7.0 / 9.6 / 80 for the linear canceller alone.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add nn_suppressor: header-only learned residual suppressor inference
93e70a7 
include/mutap/nn_suppressor.h: dense64-tanh -> GRU96 -> dense22-sigmoid
band-gain suppressor with the classical residual_suppressor's
process_block(e, yhat, out) contract — sqrt-Hann 128/64 analysis over
[previous block, current block], 22 ERB-band features of E and Yhat,
per-bin gain application on the E spectrum, one block of added latency,
noexcept/allocation-free after construction. Yhat feeds only the
features, never the signal path, so a wrong prediction can only
mis-gain a band, not synthesize signal.

tools/ml/nn_infer.cpp + test_parity.py pin the C++ against the numpy
reference on identical frames (measured 2.3e-8 relative — FFT-backend
depth); features.py's framing is now streaming-consistent (zero
prehistory warm-up frame) so the GRU state trajectories match exactly.
tests/test_nn_suppressor.cpp pins the plumbing structurally: forced
unit gains reconstruct as a transparent one-block delay at float64
depth, forced zero gains silence the output, wrong geometry throws.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Trainer: checkpoint best weights every improving epoch
aeb4ec5 
An interrupted run keeps its best-so-far model instead of losing hours
of CPU training; also detach the loss before accumulating the epoch
statistic.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add trained v1 suppressor weights and measured hybrid results
0d808a8 
pretrained/suppressor_v1.munn: 51,670 parameters, trained 12 epochs on
400 Mini-LibriSpeech mixtures via the tools/ml pipeline; every input is
MIT / CC BY 4.0 / synthesized here, so the weights are clean to ship.
C++ inference verified against the numpy reference with these exact
weights (1.0e-7 relative).

Measured (speech scenarios): the hybrid beats the classical chain's
single-talk ERLE by ~5 dB at equal near-end transparency; double-talk
suppression stays at linear-canceller level. Off-domain (the rig's
synthetic materials): degrades gracefully — near end survives at
26-39 dB SDR where DTLN-aec annihilates it (0-5 dB) — but its
double-talk contribution evaporates, so the classical coherence
suppressor remains the right default engine. README carries the full
tables and the ranked next steps.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add ml_aec_comparison notebook: classical vs hybrid vs end-to-end
e26ac7f 
notebooks/ml_aec_comparison.ipynb (built by tools/ml/build_ml_notebook.py,
same convention as the other notebooks: the builder is the source, the
.ipynb a build product with outputs committed). Two scenarios through
one meter — the rig's synthetic music double-talk and a LibriSpeech
speech scenario — with residual-level timelines per system, the metric
tables, and heatmaps of the hybrid's learned band gains in and out of
its training domain (where the coherence-gate behavior it learned, and
the off-domain flicker that costs it double-talk suppression, are both
directly visible). nn.py gains a MUNN loader so the committed
pretrained weights drive the notebook without training artifacts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Notebook: add TCLw-style coupling-loss timelines
9cb1585 
The industry chart for this experiment — send level relative to the
far end over time, with the -30 dB post-double-talk target line — for
all four systems on both scenarios. The far-end reference is local
(smoothed block power) with an activity gate, so real speech's pauses
read as gaps instead of fake infinite loss; the accompanying text
explains the meter's deliberate blindness (near-end speech counts
against 'loss' during double-talk, so a trace that stays flat there is
suppressing the talker).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add WebRTC AEC3 as a benchmark baseline
b7c73ce 
tools/ml/webrtc_aec3_infer.cpp: offline driver for the AEC3 canceller
from freedesktop's webrtc-audio-processing (BSD-3, >= 1.0 — the distro
0.3.x packages carry only the legacy canceller and are not accepted by
the CMake probe). Optional target, appears when pkg-config finds the
library; wired into run_benchmark.py (--webrtc-aec3-bin) and the
comparison notebook, build recipe in the README.

Measured on the speech scenarios (medians): 37.9 dB single-talk ERLE /
2.7 dB double-talk suppression / 39.2 dB recovery ERLE / 4.5 dB
near-end SDR — strong echo removal, and the worst near-end transparency
of the field: AEC3's suppressor clamps the talker, a telephony-shaped
trade and the opposite of what program material wants.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Notebook: rebuild with WebRTC AEC3 on every chart and table
15e7924 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add head-to-head notebook builder: AEC3 vs chain vs hybrid
97125b6 
Focused three-way comparison (the clear-verdict companion to
ml_aec_comparison): coupling-loss timelines in the industry chart idiom
with the double-talk window shaded, true-residual timelines from the
simulator's ground truth, and a four-panel scorecard over three speech
scenarios and the three synthetic materials. Fixed colorblind-validated
color per system across every figure. The executed notebook follows in
the next commit.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add executed head-to-head notebook: AEC3 vs chain vs hybrid
13359fd 
Full run committed with outputs: coupling-loss and true-residual
timelines for a speech and a music scenario, and the scorecard over
three speech scenarios / three synthetic materials. Headline medians —
speech: AEC3 37.9 dB stERLE but 4.5 dB near-end SDR; chain 30.9 /
38.2; hybrid 36.0 / 37.6. Synthetic: AEC3 degrades everywhere (26 dB
ERLE, 8.7 dB near-end SDR); chain 41.9 / 36.1 with 15 dB double-talk
suppression; hybrid 43.6 / 33.2 with the known off-domain double-talk
gap.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Head-to-head builder: add DTLN-aec 512 as fourth system
d2f7ef9 
Fourth categorical color (yellow, validated colorblind-safe in order
with the existing three), intro/verdict updated with DTLN's two faces —
best-in-field single-talk ERLE on speech, talker deletion off-domain,
10.4M parameters vs the hybrid's 51k, benchmark-only weights. The
re-executed notebook follows in the next commit.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Re-execute head-to-head notebook with DTLN-aec 512
cc0a50e 
Four systems on every figure and table. Headline medians — speech:
DTLN leads single-talk ERLE (43.4 dB) but trails on recovery (26.3)
and near-end SDR (29.3); the chain keeps the double-talk crown
(17.2 dB) and the hybrid the fidelity/ERLE balance (36.0 / 37.6).
Synthetic: both pure-ML systems lose their double-talk contribution,
but where the hybrid's talker survives at 33.2 dB SDR, DTLN's is
deleted (0.1 dB).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
aec_chain: make the post stage a template parameter
0498b4f 
aec_chain<Sample, Canceller, Post = residual_suppressor<Sample>> — the
post-filter engine becomes pluggable with the same capability-detection
idiom the canceller side already uses: any type with the suppressor
contract slots in, and the receive guard / re-convergence rescue key on
echo_explained() when the post provides it (a post without the
statistic certifies the guard immediately and disables the rescue,
via if constexpr (requires ...)).

The classical path is provably untouched: aec_chain<double> still
denotes raw FD-Kalman + residual_suppressor, and the branchless-parity
fingerprint (tests/branchless_parity_check.cpp, 600 preset blocks) is
bit-identical before and after this change for both suppressor forms
(d515415bee83b420). ItuChain/postfilter suites pass unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
nn_suppressor v2: self-describing geometry + full chain contract
e18bc33 
The learned suppressor now composes as aec_chain's Post engine on any
trained geometry, not just the 16 kHz/hop-64 benchmark model:

- nn_geometry (sample_rate, hop, bands, dense, gru) travels WITH the
  weights: MUNN0002 files carry it in a header; legacy MUNN0001 files
  imply the original geometry. Band matrix, buffers and the network all
  size themselves from it; the chain's block size must equal the
  trained hop (validated).
- Config-only construction (weights ride in the config) so
  aec_chain<Sample, Canceller, nn_suppressor<Sample>> builds it like
  any other post stage.
- echo_explained(): the classical guard/rescue statistic (smoothed
  Yhat power over mic power, mic reconstructed as E + Yhat per bin) so
  the receive guard and re-convergence rescue work unchanged.
- Comfort noise: two-window minimum-statistics floor tracker on the
  per-bin E PSD with deterministic-reset xorshift fill, same contract
  and defaults as the classical engine.

Exporter writes MUNN0002 (geometry from the trainer npz when present);
the Python reference loads both formats and sizes its GRU from the
weights. Tests: transparency/silence/comfort-floor/echo_explained/
geometry-rejection/chain-composition all pass; Python-C++ parity
unchanged at 2.3e-8 relative; legacy MUNN0001 verified end-to-end.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add aec_chain_nn preset, C ABI nn engine, and 48 kHz training pipeline
dc7c3f1 
include/mutap/nn_chain.h: aec_chain_nn<Sample> (raw FD-Kalman +
nn_suppressor) and aec_chain_nn_preset — the classical preset's
canceller/guard/rescue/shadow calibration verbatim, with the learned
post engine's floor-tracker constants rescaled by the same geometry
ratio. tools/capi: MutapAec becomes the repo's std::variant idiom;
mutap_aec_create_nn(weights_path) constructs the learned-engine chain;
every entry point dispatches via std::visit. mutap_ffi.AecChain grows
nn_weights so the benchmark measures the real C++ chain.

tools/ml geometry parameterization for the 48 kHz-native model:
features.Geometry (GEOM16 benchmark / GEOM48 = mutap.aec~'s block-256
default, 26 bands), make_dataset --rate/--taps/--near-materials with
numpy ports of the rig's synthetic near-end materials (frequencies
pinned in Hz) mixed into training — the measured fix for the hybrid's
off-domain double-talk gap — spectral 3x upsampling for LibriSpeech,
and FFT convolution; the trainer sizes the model from the shards'
geometry and stamps it into the checkpoint.

Sanity (16 kHz v1 weights through the C ABI on a speech scenario):
chain-nn 58.1 dB stERLE / 37.8 dB neSDR vs classic 50.9 / 48.6, with
the known v1 double-talk gap (9.6 vs 17.2 dB) the 48 kHz retrain's
mixed materials target. Parity 2.3e-8 unchanged; NnSuppressor suite
passes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Add the 48 kHz-native trained model (v2) and the embedded-weights path
0c0b7ff 
pretrained/suppressor_v2_48k.munn: MUNN0002, 52,570 parameters, trained
12 epochs on 400 mixtures at the mutap.aec~ geometry (48 kHz, hop 256,
26 bands) with the off-domain fix in the data: near ends drawn from
speech AND the rig's synthetic families (AR envelope, voiced, music).
Every input remains clean-licensed (LibriSpeech CC BY 4.0 upsampled,
synthetic material ours).

Measured through the C ABI chain at 48 kHz (3 speech scenarios,
medians): nn48 53.4 dB single-talk ERLE / 11.2 dB double-talk
suppression / 34.3 dB near-end SDR vs classic 34.5 / 12.6 / 30.3 — the
learned engine now leads ERLE and transparency at near-parity
double-talk. Off-domain (music double-talk): 4.0 dB suppression at
30.2 dB near-end SDR, vs the 16 kHz v1's ~0 dB collapse — the training
mix closed most of the gap; classic still leads there (9.1 dB) and
remains the default.

nn_suppressor gains parse_nn_suppressor_weights (in-memory MUNN image)
so a package can embed the default model as a byte array;
export_weights.py --header emits that array.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
tools/ml README: v2 48 kHz model results and shipping surface
26f12a3 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Docs: learned engine in the book chapter, README, and HANDOFF
ee21466 
book/src/echo-cancellation.md gains 'The learned engine: @PostFilter 2'
— what it is (a 52k-parameter band-gain network in the post-filter
slot, structurally unable to synthesize signal), the measured trade
(~53 vs ~35 dB single-talk ERLE at better transparency; classical keeps
the double-talk crown off-domain and stays the default), and the
practicalities (@block follows the model, @model loads user-trained
weights, all training inputs clean-licensed) — plus the revised knobs
list. README gains the library-status bullet for nn_suppressor /
aec_chain_nn / tools/ml; HANDOFF records the whole effort as Stage 6
with the filed follow-ups (coherence feature, int8/CMSIS-NN, 44.1 kHz
model, ITU characterization run).

webrtc_aec3_infer gains an optional sample-rate argument so the AEC3
baseline joins the notebooks' 48 kHz shipping-geometry section.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Survey notebook: pointer to the head-to-head and the shipping v2 model
08361e4 
Markdown-only cross-reference (no outputs touched), synchronized
between the builder and the committed notebook.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Head-to-head notebook: the shipping geometry (48 kHz, block 256)
9cf1ea3 
New closing section measured at mutap.aec~'s native geometry through
the engines' real deployment paths (the C ABI chain — exactly what
@PostFilter 1/2 run): the v2 learned engine posts 53.4 dB single-talk
ERLE / 51.9 recovery / 34.3 near-end SDR vs the classical chain's
34.5 / 31.2 / 30.3, with double-talk suppression within 1.5 dB
(11.2 vs 12.6); WebRTC AEC3 at 48 kHz shows its telephony trade again
(42.3 dB ERLE, 7.8 dB near-end SDR). DTLN-aec sits this section out —
its models are inherently 16 kHz. webrtc_aec3_infer gained the
sample-rate argument this section uses.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
Head-to-head notebook: rewrite the narrative for the whole scope
3a61cd9 
The prose had accreted section by section (three systems, then four,
then the 48 kHz addendum) and read like it. Rewritten as one arc:

- The intro announces the two-part structure and why it exists (DTLN
  pins Part 1 to the 16 kHz benchmark geometry; Part 2 is the 48 kHz
  shipping geometry without it), and that 'MuTap learned' appears in
  two generations — the v1 prototype in Part 1, the shipping v2 model
  in Part 2.
- One name and one color per engine identity everywhere: 'MuTap
  classical' / 'MuTap learned (v1|v2)' in prose, legends and tables
  alike (the chain/hybrid drift is gone).
- Section transitions written for the arc: the coupling-loss reading
  interprets both charts together, the residual view is introduced as
  what the deployable meter cannot see, Part 1 closes with a
  per-system summary that sets up Part 2, and Part 2 states its
  realism deltas (real C ABI deployment path, AEC3 at 48 kHz)
  explicitly.
- The verdict now accounts for everything measured, including the
  v1 -> v2 off-domain repair and the licensing asymmetry.

Measurements unchanged (deterministic seeds; tables verified identical
to the previous execution). Re-executed and committed with outputs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@claude
export_weights --header: clang-format guards on the generated array
70875ce 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
@tap Claude

tap commented Aug 8, 2026

Copy link
Copy Markdown
Owner Author

CI status note: the macOS AppleClang failure (itu_echo.EchoStability<float>, 1 of 187) is pre-existing on the base branch — main's most recent CI run, on commit 41b9ef7 (this branch's exact merge base), fails the same single test in the same job, and main's macOS history alternates green/red on it. Every check this PR can affect is green here (Linux GCC full suite 187/187 in-run; the branchless-parity fingerprint pins the classical chain bit-identical), so I'm not chasing the macOS float32 flake in this PR. It looks like vDSP-backend float32 numerics sitting on a threshold — worth its own issue if it keeps recurring on main.

Generated by Claude Code

@tap tap mentioned this pull request Aug 8, 2026
Closed
@tap
tap merged commit 1d258d3 into main Aug 8, 2026
26 of 28 checks passed
tap pushed a commit to tap/MuTap-Max that referenced this pull request Aug 8, 2026
@claude
Repoint MuTap submodule at the merged tree on main
c3f0bab 
tap/MuTap#33 merged (rebase); the pin moves from the branch head to the
identical tree on main (1d258d3), per the submodule rule in the PR
notes. Package builds and tests green against the pinned submodule
itself (no MuTap_ROOT override).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
tap pushed a commit to tap/MuTap-Max that referenced this pull request Aug 10, 2026
@claude @tap
Repoint MuTap submodule at the merged tree on main
eef3143 
tap/MuTap#33 merged (rebase); the pin moves from the branch head to the
identical tree on main (1d258d3), per the submodule rule in the PR
notes. Package builds and tests green against the pinned submodule
itself (no MuTap_ROOT override).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y3Y23Xavpd31igC1mDDcWg
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@tap @claude