commit 94b82c1b15eb59e8bc223e6ba18e209a0fe67236
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Wed Jul 29 11:54:09 2026 -0700

    V2 rx: enable coarse CP timing adjustment by default
    
    rx->timing_adj gated adjust_timing() but was never set anywhere in the
    C port (rade_rx_v2_init, rade_api.c, rade_rx_wav.c), leaving the
    coarse +/-40 sample CP-centering correction dead. The receiver would
    lock onto whatever sub-symbol sample-phase timing offset the CP
    autocorrelation peak happened to land on at acquisition and keep it
    for the rest of the session, causing loss to vary by up to +15%
    depending on signal start-of-burst alignment (verified via a sweep of
    1-40ms of silence prepended before TX, period exactly matches
    RADE_V2_SYM_LEN = 160 samples = 20ms).
    
    The Python reference (rx2.py) enables this almost immediately by
    default via --timing_adj_at 0 (receiver.timing_adj = 1 once s > 0).
    Mirror that by enabling it unconditionally in rade_rx_v2_init().

commit ffd18cde1bf4cf877dcec82c73324924b6953d4c
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 28 19:25:15 2026 -0700

    Pass --enable-dnn-debug-float to Opus's configure instead of undef'ing it.
    
    Opus's own configure disables DISABLE_DEBUG_FLOAT... err, disables the
    float-precision debug weights (defines DISABLE_DEBUG_FLOAT) unless
    --enable-dnn-debug-float is passed. Passing that flag directly to Opus's
    ./configure invocation means the macro is never defined in config.h in the
    first place, so rade_fargan_config.h no longer needs to #undef it after
    the fact. Verified rade_c_v2_tx/enc/dec/text/wav/wav_real all still pass,
    and the Opus SSE/ARM build warnings this branch silences stay silenced.
    
    Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
    Claude-Session: https://claude.ai/code/session_011vRA8TJ4naGLX86tmJwULM

commit 1e9c1210f21f71743f2d29e885343e5bddb98db9
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 28 19:10:53 2026 -0700

    Preserve float-precision DNN debug weights when silencing Opus warnings.
    
    Including Opus's generated config.h (to silence build warnings) also pulls
    in DISABLE_DEBUG_FLOAT, which strips the float-precision copies of the V2
    encoder/decoder weights out of rade_enc_v2_data.c/rade_dec_v2_data.c,
    leaving only the lower-precision int8-quantized weights. This pushed
    rade_c_v2_tx's diff_rms/sig_rms from 0.0124 to 0.0348, over the 0.03
    threshold. Undefine DISABLE_DEBUG_FLOAT after including Opus's config.h so
    RADE keeps matching the Python/PyTorch float32 reference.
    
    Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
    Claude-Session: https://claude.ai/code/session_011vRA8TJ4naGLX86tmJwULM

commit 87b3e8f766b2a84bf5a0513163129e6122d528ee
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 28 15:14:24 2026 -0700

    Keep FARGAN config file include out of .h files.

commit fb6a1f2b81e93b453867e09879722ae4a1171906
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 28 15:02:40 2026 -0700

    Add missed file.

commit e05d5ba525004789aa55401294b30383cc85cf50
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 28 14:58:14 2026 -0700

    Silence Opus-related build/link warnings.

commit 5eb39560c64ffc080518966ac2a7f0eceb73c9be
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Sat Jul 18 00:17:50 2026 -0700

    Add feature file output capability to rade_rx_wav/rade_tx_wav.

commit 7503832b378e4d442c555e6dc5d9833950093614
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Sat Jul 18 00:17:50 2026 -0700

    Add feature file output capability to rade_rx_wav/rade_tx_wav.

commit 22d6e84f206a97b9b376d42a16eaf36b91d8b52c
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Tue Jul 14 15:18:00 2026 -0700

    Cap raw SNR at <= 0.1, not 0.0.

commit 844047ad8eb853719213b5615e95de29c96ecad0
Author: Mooneer Salem <mooneer@gmail.com>
Date:   Fri Jul 10 10:21:39 2026 -0700

    Update loss.py to match Python RADEV2 repo.