spectral_reverb returns under the ABI rules, measured against the plates (phase 2 item 7, part 2)

[tap]

Phase 2 item 7, part 2 (HANDOFF item 11). tap::mu::spectral_reverb, the per-bin reverb from the karaoke branch (8abe958), comes back under the ABI rules and is measured on part 1's harness beside the as-shipped Dattorro plate.

Verdict: it loses to the plate everywhere measured. It stays in the repo only as a measured negative result, and its header says so. Removing it again is the alternative; I recommend against it, because the gated rows that hold the verdict need the class.

What changed

• include/mutap/spectral_reverb.h:
  • defined inside tap::mu::inline TAP_DSP_FFT_ABI;
  • the 2 × block FFT size goes through fft_detail::checked_fft_size before any buffer is sized (working note 6 (b): on the M55 an unchecked size is a release-mode HardFault);
  • an afc_stage (process_block(in, out, n) noexcept, n a multiple of the block), allocating only at construction, latency() one block;
  • spectral_shaping::flat / from_path with an explicit shape_max;
  • shaped from F̂ (reshape_from_impulse_response) or from the coherent bus sum |Σ F̂_m| (reshape_from_impulse_responses).
• tests/test_fft_engine_contract.cpp pins it as the sixth embedder: the tag by qualified name and typeid, and a size-gate sweep in both profiles plus the explicit CMSIS rows. mutap/fft.h comments go from "five" to "six".
• tests/test_spectral_reverb.cpp covers the float/double plumbing and runs in both emulated selections (bare_metal_main.cpp and TEST_FILTER, kept in sync; the on-target count comment is now 118).
• decorrelated_loop's forward stage now takes a stage's latency() out of the forward delay. The plates report 0, so their rows repeat: the S1 cost medians match part 1's grid to the printed digit.
• New harness pieces:
  • tests/support/spectral_rig.h: the measurement and the per-bin hypothesis statistics;
  • gated host rows tests/test_spectral_reverb_host.cpp;
  • the MUTAP_SLOW sweep tests/test_spectral_reverb_sweep.cpp (ProbeConvergence, Grid);
  • karaoke_ramp_dump --reverb spectral and reverb_audible.py --plate spectral.
• Docs: docs/reverb-afc.md gets a "Spectral reverb" section, and the headline, intro, not-run list and provenance are updated. docs: the S3 wet-0.30 audible rows examined — bounded lines at the plate's own modes, loop-dependent #84's S3 paragraphs are kept. HANDOFF item 11 is rewritten as resolved; items 13 and working note 6 are updated.
• Follow-up commit d912973:
  • the 80 s host rows move behind MUTAP_SLOW, and 20 s proxies run by default;
  • the doc says, where the plate rows are described, that a stage's latency is budgeted inside the loop delay;
  • the S3 chain-output flags become an open item next to the plates' ‡ lines.

Numbers

All from macOS 15.7 x86_64, AppleClang 17, Release, five seed sets.

Probe convergence (S1, wet 0.30, rt60 1 s; median ASG re the dry open loop). Unlike the plates (10 s), only the cabin's flat, shape_max 1 and 2 held rows converged, at 80 s. shape_max 4 and every mt5 chain row had not converged by 160 s.

Row	10 s	20 s	40 s	80 s	160 s	Converged
cabin, held, canceller alone	+11.66	+11.74	+11.74	+11.74		10 s
cabin, held, flat	−3.87	−4.67	−5.64	−6.62	−6.82	80 s
cabin, held, shape_max 1	+3.36	+2.27	+1.48	−0.37	−0.27	80 s
cabin, held, shape_max 2	+5.41	+3.63	+3.63	+2.27	+2.36	80 s
cabin, held, shape_max 4	−2.01	−3.30	−5.74	−17.46	−19.02	not
mt5, held, flat	+1.21	+1.00	+0.41	−0.27	−0.57	not
mt5, held, shape_max 1	+10.57	+8.03	+9.49	+8.03	+8.42	not
cabin, held, flat, no canceller	−29.14	−29.75	−30.14	−30.33	−30.43	80 s

The doc has the full table, mt5 and the speech envelope included.

Behind the canceller. Six rooms, held note, 40 s probes, so not converged; the doc explains why this mostly understates the cost. Values are the median of room medians.

	Plate 0.5 cost	Spectral cost (flat / sm1 / sm2 / sm4)
S1, w 0.15 (rt60 1.18)	+0.49	+5.66 / +2.83 / +7.03 / +14.26
S1, w 0.30 (rt60 1.18)	+0.88	+13.09 / +10.74 / +13.48 / +30.57
S3, w 0.15 (rt60 1.18)	−0.59	+10.16 / +10.94 / +11.82 / +15.04
S3, w 0.30 (rt60 1.18)	−1.37	+18.75 / +15.14 / +16.70 / +15.23

• Paired by seed, no room's median favours the spectral reverb for 22 of 24 rows; for the other two (flat and shape_max 1 at S1, w 0.15), 1 of 6 rooms does.
• At equal wet the spectral stage is 4.30 to 14.75 dB louder than the plate's mix.
• Speech envelope (cabin and mt5, S1, w 0.30, 80 s): it costs +4.30 to +17.09 dB where the plate raised the limit.

Open loop (no canceller, 160 s): flat sits 17.15 to 31.80 dB below the dry room (room medians), confirming HANDOFF 11's −18 to −25. Shaped rows sit 12.56 to 29.84 dB below.

Shaped vs flat:

• shape_max 1 is +1.17 to +4.30 over flat at S1.
• shape_max 4 is −7.91 to −12.79 at S1, in every room.
• At S3, w 0.30, every ceiling is +0.98 to +3.71.

Hypothesis (Spearman over the 65 bins, at the moment of shaping, 30 runs per row):

• ρ(decay time, |F − F̂|) is −0.336 to −0.688. The long decays sit where the absolute misalignment is small.
• The worst-quarter energy share is 0.022 to 0.178, against flat's 0.246.
• ρ against the relative misalignment |F − F̂|/|F| is +0.161 to +0.526.
• The residual-loop indicator (max_k |F − F̂|·g_k against flat) is −0.85 to −3.25 dB at shape_max 1 and +5.49 to +9.72 at shape_max 4.
• None of these predicts a run's own shaped − flat result: ρ over 30 runs ranges −0.351 to +0.699, with no consistent sign.

Audible (cabin, ramp + criterion, PROTOCOL §7.3, rt60 1.1846 s; chain output, with c in brackets):

	S1 aud	S1 run	S3 aud	S3 run
plate 0.5, w 0.15 (part 1)	+16.53	+21.46	+10.38	+23.01
plate 0.5, w 0.30 (part 1)	+12.89	+22.07	−4.77 ‡	+23.31
flat, w 0.15	−7.42 (−6.74)	+4.73	−17.78 (−2.05)	+11.66
shape_max 1, w 0.15	+11.90 (+11.90)	+18.31	−13.14 (+4.98)	+17.98
shape_max 4, w 0.15	−9.40 (−7.99)	+1.16	−17.23 (+8.34)	+14.11
flat, w 0.30	−6.30 (−6.42)	−2.82	−12.77 (−9.24)	+3.56

Every spectral row is audible earlier than the plate at the same wet. The doc has all 16 rows.

Gated rows. Converged probes run behind MUTAP_SLOW. The default run carries 20 s, two-seed-set proxies with the same directions. Short probes understate both the spectral cost and the open loop's danger, so a pass is conservative.

Row	Run	Probe, sets	Gate	Measured	Time (3 threads)
DecayAndLevelBesideThePlate	default	analytic	T30 = rt60 ± 0.005 s; level > 0 dB	T30 1.0000 / 1.1846 s; +3.40 … +8.22 dB	0.25 s
UnsafeWithoutTheCancellerShortProbe	default	20 s, 2	flat mean < −15; shape_max 1 mean < −8	cabin −29.80 / −19.79; mt5 −22.42 / −13.98	27.03 s
CostsMoreThanThePlateShortProbe	default	20 s, 2	plate − spectral: flat mean > +6; shape_max 1 mean > +3	+14.66 / +9.78 (per set: +15.75 +13.56 / +11.62 +7.94)	77.25 s
UnsafeWithoutTheCanceller	MUTAP_SLOW	80 s, 5	flat median < −15; shape_max 1 median < −8	cabin −30.51 / −19.61; mt5 −22.77 / −14.69	189.41 s
CostsMoreThanThePlateBehindTheCanceller	MUTAP_SLOW	80 s, 5	cost vs canceller alone: flat > +8; shape_max 1 > +4	+17.19 / +8.44	696.95 s

The 20 s values are recorded beside the 80 s ones in the test comments, for this host only so far; CI's legs run the default rows.

Verification

• cmake -DCMAKE_BUILD_TYPE=Release -DMUTAP_WERROR=ON -DMUTAP_BUILD_KARAOKE_DUMP=ON: built clean.
• Full ctest, default suite, at d912973: 420/420 passed, 3678.66 s serial with MUTAP_SLOW_THREADS=3. At aadf7db, with the 80 s rows still in the default run, it was 418/418 in 4479.84 s.
  • The spectral default rows took 99.92 s inside ctest: 0.24 s, 26.44 s and 73.24 s.
  • The two slow host rows and the two sweep entries report Skipped.
  • mutap_fingerprint passed.
• pre-commit via xargs: passed.
• scripts/tidy.sh on the 8 changed host TUs: clean. A direct clang-tidy-18 -p build-tidy count: 0 project warnings.
• Fingerprints and icount are unchanged: the new class is in neither the fingerprint harness nor any icount workload, and fft.h changed in comments only.

Sweep wall times, on 3 threads of the shared Intel Mac:

Run	Wall time
ProbeConvergence, held	11349.4 s
ProbeConvergence, speech	6849.3 s
Grid, held	35418.6 s
Audible	1380.5 s
UnsafeWithoutTheCanceller	189.41 s
CostsMoreThanThePlateBehindTheCanceller	696.95 s

Not run (the budget)

• The six-room speech grid; speech is measured in the two ProbeConvergence rooms only.
• Converged probes for the grid.
• S3 on speech.
• float32 in the loop.
• Two microphones: the bus-sum shaping has plumbing tests only.
• The backing track.
• Audible rows beyond the cabin held note at rt60 1.1846 s.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A