Fix filter test/doc claims: retracted DF1 limit-cycle rationale, notch-depth overreach, stale drive-branch wording

src/core/instrument/engine/filter/CLAUDE.md

@@ -90,10 +90,15 @@ persisted field lands on it rather than on the SEM leg.
   and LP **together** across the whole sweep, `bp == 0` throughout. The notch is not tuned in:
   HP and LP sit at exactly +90° and −90° at the corner, so equal weights cancel there by
   construction. Here the corner magnitude deliberately goes to **zero** at the centre —
-  measured worst case −88 dB across every rate/cutoff/Q, typically −110 to −145 dB. The fold
-  makes that structural rather than a runtime near-miss: `m2 = lp - hp` is **exactly** `0.0f`
-  at the centre, because `cos` and `sin` of π/4 differ by about an ulp of *double*, nine
-  orders below float's spacing there, so they narrow to one float.
+  measured worst case −88 dB on the shipped `{250, 1000, 4000}` Hz cutoff grid, typically −110 to
+  −145 dB. Over the full control range (20 Hz – 20 kHz, Q 0.1 – 10) the worst residual is
+  shallower — −69.8 dB at 192 kHz / 30 Hz / Q=10 — from float conditioning in the folded
+  `x − k·v1` term as `fc/sr → 1e-4` at high Q; it is Q-dependent (Q=0.1 holds −110 dB everywhere)
+  and still an excellent notch, not a broadband defect. `test_filter.cpp`'s null test covers this
+  full range with a Q-scaled threshold rather than the flat −74 dB the shipped grid alone would
+  justify. The fold makes the centre's cancellation structural rather than a runtime near-miss:
+  `m2 = lp - hp` is **exactly** `0.0f` at the centre, because `cos` and `sin` of π/4 differ by
+  about an ulp of *double*, nine orders below float's spacing there, so they narrow to one float.
 
 SEM's zero is at the **notch frequency**, not a broadband level sag — off the corner the pair
 is still equal-power, so neither law's legs dip. Measuring that requires dividing by each

src/core/instrument/engine/filter/filter_saturate.h

@@ -11,7 +11,9 @@ namespace reasampler::instrument::engine::filter {
 //
 // Three properties are load-bearing and none of them are tuning:
 //   - depth == 0 makes this ALGEBRAICALLY the identity (x / sqrt(1) == x, exact in IEEE), so
-//     drive = 0 is bit-exact linear with no branch and no special case on the hot path.
+//     drive = 0 is bit-exact linear whether or not the caller special-cases it. (voice_filter.h
+//     gates the call on drive != 0 anyway, but as a perf optimization, not because correctness
+//     needs it.)
 //   - |softLimit(x, d)| <= |x| for every d, so dropping it into the resonance state update can
 //     only ever shrink the state. The filter therefore cannot gain energy from the drive stage:
 //     stability at any Q and any cutoff is structural, not a tuned margin, and it can never