cost-model: model packed-A re-stream cost

[kali]

The per-kernel a·padded_work + b·n_tiles + c cannot express re-streaming a large weight once per n-pass: invisible when a kernel is timed in isolation (weight cache-resident) but real in a model (weight evicted between layers), where a wide-nr kernel that re-streams A fewer times wins a tie a narrower one appears to win in isolation. Add a per-model restream coefficient, calibrated from a cold-cache probe (regen --cold-probe/--cold-dataset), left 0 until calibrated. Re-gather intel_avx512 and calibrate it.

[github-actions]

🔴 Bench vs main — 1 speed regression(s) · ⚠️ 12 secondary

Reference: 2026-07-16 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +5.4%	mobilenet_v1_1 evaltime · pass	apple-m1-max	18.7 ms → 19.7 ms

Improvements

Δ	metric	device	main → PR
🟢 -43.5%	dummy_conmer_12M evaltime · pulse_120ms	apple-m1-max	3.22 ms → 1.82 ms
🟢 -43.1%	parakeet_tdt_600m_v3_f32f32_encoder_1s evaltime · cpu	apple-m1-max	217 ms → 123 ms
🟢 -38.4%	en_tdnn_15M evaltime · pulse_120ms	apple-m1-max	3.77 ms → 2.32 ms
🟢 -35.7%	en_tdnn_8M evaltime · pulse_120ms	apple-m1-max	1.82 ms → 1.17 ms
🟢 -30.8%	hey_snips_v4_model17 evaltime · 2sec	apple-m1-max	1.42 ms → 0.986 ms

+24 more improvement(s)

Δ	metric	device	main → PR
🟢 -24.9%	voicecom_float evaltime · 2sec	cortex-a9	235 ms → 177 ms
🟢 -24.1%	voicecom_fake_quant evaltime · 2sec	cortex-a9	259 ms → 197 ms
🟢 -20.4%	en_tdnn_8M evaltime · pulse_180ms	apple-m1-max	2.17 ms → 1.73 ms
🟢 -19.3%	hey_snips_v31 evaltime · 400ms	apple-m1-max	0.115 ms → 0.0932 ms
🟢 -19.2%	parakeet_tdt_600m_v3_f32f32_encoder_1s evaltime · cpu	i9-11900kb_rtx-4060	293 ms → 237 ms
🟢 -19.1%	trunet evaltime · pulse1_f32	apple-m1-max	0.204 ms → 0.165 ms
🟢 -17.6%	en_tdnn_15M_nnef evaltime · pulse_240ms	apple-m1-max	4.7 ms → 3.87 ms
🟢 -17.2%	en_tdnn_15M evaltime · pulse_240ms	apple-m1-max	4.84 ms → 4.01 ms
🟢 -15.6%	en_tdnn_pyt_15M evaltime · pulse_120ms	apple-m1-max	4.13 ms → 3.49 ms
🟢 -12.8%	arm_ml_kws_cnn_m evaltime · pass	apple-m1-max	0.15 ms → 0.131 ms
🟢 -10.7%	hey_snips_v4_model17 evaltime · pulse8	apple-m1-max	0.135 ms → 0.121 ms
🟢 -9.6%	en_tdnn_15M evaltime · 2600ms	cortex-a9	2.78 s → 2.51 s
🟢 -9.2%	voicecom_fake_quant evaltime · 2sec	apple-m1-max	1.65 ms → 1.49 ms
🟢 -9.1%	en_tdnn_8M evaltime · 2600ms	cortex-a9	1.32 s → 1.2 s
🟢 -7.9%	en_tdnn_8M_nnef evaltime · pulse_240ms	apple-m1-max	2.29 ms → 2.11 ms
🟢 -7.8%	mdl_en_2019_Q3_librispeech_onnx evaltime · pulse_240ms	apple-m1-max	0.807 ms → 0.745 ms
🟢 -6.8%	en_tdnn_8M evaltime · pulse_240ms	apple-m1-max	2.34 ms → 2.18 ms
🟢 -6.5%	mobilenet_v1_1 evaltime · pass	cortex-a55	328 ms → 307 ms
🟢 -6.2%	en_tdnn_15M evaltime · 2600ms	cortex-a7	2.92 s → 2.74 s
🟢 -6.1%	voicecom_float evaltime · 2sec	apple-m1-max	1.39 ms → 1.3 ms
🟢 -5.7%	inceptionv3 evaltime · pass	cortex-a7	8.53 s → 8.05 s
🟢 -5.7%	inceptionv3 evaltime · pass	cortex-a9	6.95 s → 6.55 s
🟢 -5.4%	mobilenet_v1_1 evaltime · pass	cortex-a9	1.32 s → 1.25 s
🟢 -5.4%	mobilenet_v1_1 evaltime · pass	cortex-a7	1.29 s → 1.23 s

⚠️ 12 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +107.1%	voicecom_float load · 2sec	cortex-a55	155 ms → 321 ms
⚠️ +89.8%	voicecom_float load+optimize · 2sec	cortex-a55	187 ms → 355 ms
⚠️ +9.1%	trunet RSS @ load · pulse1_f32	cortex-a9	23.2 MB → 25.3 MB
⚠️ +6.6%	speaker_id RSS @ load · pulse8	cortex-a9	19.1 MB → 20.3 MB
⚠️ +4.5%	voicecom_float heap @ ready · 2sec	cortex-a55	2.88 MB → 3.01 MB
⚠️ +4.4%	voicecom_fake_quant heap @ ready · 2sec	cortex-a55	3.06 MB → 3.19 MB
⚠️ +3.2%	parakeet_tdt_600m_v3_f32f32_preprocessor_1s heap @ ready · cpu	i9-11900kb_rtx-4060	643 kB → 664 kB
⚠️ +3.1%	voicecom_float heap @ ready · 2sec	i9-11900kb_rtx-4060	3.13 MB → 3.23 MB
⚠️ +3.0%	voicecom_fake_quant heap @ ready · 2sec	i9-11900kb_rtx-4060	3.31 MB → 3.41 MB
⚠️ +2.7%	parakeet_tdt_600m_v3_f32f32_preprocessor_1s heap @ ready · cpu	apple-m1-max	646 kB → 664 kB
⚠️ +2.4%	hey_snips_v31 heap @ ready · 400ms	i9-11900kb_rtx-4060	1.03 MB → 1.05 MB
⚠️ +2.4%	trunet heap @ ready · pulse1_f32	apple-m1-max	4.1 MB → 4.2 MB