Four effects fail on the 11.0.0-alpha.2 integer mixer (audiodsp 46f78dc and later) #245

audiodsp 46f78dc took CircuitPython 11.0.0-alpha.2's integer mixer: a voice scales as (x * level) >> 15 where it used to scale by level / 32767 and truncate toward zero. The new arithmetic rounds toward minus infinity, so a voice below unity now leaves -1 where the old one left 0 on small negative samples. Four effects' tests fail from that commit on (they all pass at 1877bb9, the commit before, with the same effect code), and none of them is a one-line follow-up:

• StereoWidener. A Mixer clamps panning to 32767/32768, so a hard-panned voice's far lane is scaled by 1/32768 and reads -1 on every negative sample (before: 0). The Haas copy is no longer exactly the right channel delayed: test_the_constructor_delay_plays_exactly, H1TheCopy (two tests), H3TheGlide, test_reset_empties_the_line_and_returns_to_patch_0. A pan of 1.0 doesn't help, because of the clamp.
• Distortion. The leg mix runs below unity ahead of about 67 dB of pre-clip gain, so the -0.5 LSB mean bias comes out as standing negative DC on quiet input. At 44.1 kHz, Distortion at maximum, a 1 kHz sine at 4 LSB now renders -0.070 of full scale of DC (before: silence), and at -40 dBFS input -0.017 (before: -0.0003). The F1/F2/F4 rows, their planted-fault guards and TheResetResidue (909 against 800) go red.
• Overdrive. Patch 6 at 44.1 kHz stands 1 LSB of DC into digital silence (before: 0): test_the_planted_class_stands_dc_where_this_one_does_not.
• DeEsser. D1 got better on every patch (patch 0 from 0.386 to 0.059 dB, patch 3 from 2.262 to 0.583, patch 4 from 2.479 to 1.321). The sweep holds the recorded verdict "D1 misses at the HF-only patches", and patch 3 no longer misses.

MultiTapDelay, AirSpace and the Harmonizer tests needed only to drop or model the old arithmetic; that is done in #243, which moves the pin past 46f78dc.