v0.3: profile-ladder re-pin — economy at 18 kHz, balanced, super_economy, ladder notebook

[tap]

What this changes

The profile ladder is re-pinned as four tiers: the default economy() moves to an 18 kHz passband at 58/38 taps (−26%/−14% MACs vs the previous default), the former economy design continues bit-for-bit unchanged as balanced(), super_economy() (16 kHz, 40/28 taps) joins as the explicit voice/comms tier, and transparent() is untouched. A new committed-executed notebook (notebooks/profile_ladder.ipynb) compares all four tiers with measurements taken through the shipping engine. Version bumps to 0.3.0.

Why

The "economy18" spec-relaxation experiment showed the 19→18 kHz passband move is the same species of inaudible speed-first trade that put economy at 19 kHz rather than transparent's 20 — and it buys a quarter off the expensive direction's hot loop on every embedded target (measured: M33 −23.5/−25.5% down q15/float, M55 −21.4/−22.6%, Hexagon −21.8/−23.3%). balanced preserves the old default for content that needs the 18–19 kHz shelf flat. super_economy makes a different promise — an audibly shelved top octave for half of balanced's compute — so it is never a default, opt-in by name only, positioned for voice/comms/Bluetooth-class links; in Q15 it is quieter than transparent at a fifth of the taps.

Tap counts are the minimal even counts meeting the 70 dB stopband with ≥ 1 dB margin on a fine (12.5 Hz) sweep grid; one quirk is recorded in PLAN §4 (up-direction 40 taps fails while 38 passes — Kaiser sidelobe peaking is non-monotonic near threshold).

Verification

• Host suite 82/82 green (Linux/GCC Release, -Werror), clang-tidy clean on every changed TU; CI matrix on the branch: linux/macOS/Windows, ASan+UBSan, Cortex-M33/M55 and Hexagon emulated legs all green.
• All three test legs re-measured, not just renamed: scipy reference vectors regenerated for the eight direction × profile cases; cross-validation floors re-pinned at 1.2e-5 down / 3.1e-5 up (−98/−90 dB), still equal at async L=512 and L=1024 (test_cross_validation.cpp); Q15 flagship 76.5 dB, 997 Hz float imaging floor 91.3 dB (test_converter_fixed_point.cpp / test_converter.cpp [ measured ] lines).
• Instruction-count ratchet: baselines re-recorded from the CI run's own deterministic measurements on all three targets (30 scenarios, the designed serial-harvest path), including first baselines for the two new up_q15_se/down_q15_se scenarios; the transparent legs moved ≤ ±1.3%. The harvest-commit CI run re-gates everything at the new values.
• bluetooth_bridge locks and recovers at 72.8 dB SNR (> 70 gate), total latency 1.93 ms.
• Margin notes stated plainly: the 23 kHz stopband-tone folded-alias margin thins from 12.9 to 2.9 dB over the ≥ 71 dB contract line (worst audible-band product improves to −96 dBFS), and the bridge SNR margin thins from ~6 to 2.8 dB.

Notes for the reviewer

• Contract change. The default profile's output changes for every consumer that constructs a converter without naming a profile — different filter, different latency (latency_frames() 39→29 down, 22→19 up), different (better) instruction counts. balanced() is the bit-exact escape hatch to the v0.2 default. The C ABI grows tags 2 = balanced, 3 = super_economy; tags 0/1 keep their meanings (0 = economy now denotes the new design).
• Notebooks re-executed, because behavior changed: design_spike.ipynb (all eight designs, minimality asserts included), ratio_demo.ipynb (18 kHz edge asserts), and the new profile_ladder.ipynb — all committed executed.

---

Generated by Claude Code