# THE SYSTEM-libm MANIFEST — every call to a system transcendental OUTSIDE the deterministic zone.
# Checked by tools/lint/check-det-math.mjs; read that file first for what the zone is and why this list
# is per-FILE rather than per-call-site.
#
# Format:   <path>  [<fn>*<count> ...]  <disposition>  <reason>
#
# The bracketed multiset is the exact set of governed functions the file calls. It is what makes this a
# gate rather than a comment: a call added, removed, or swapped for a different function moves the line
# and the build goes red until somebody re-reads the file. A line move or a whitespace change does not.
#
# Dispositions, and what each one CLAIMS:
#   retain-rt        the result is a shipped product's audio. Changing it changes TabbyEQ's or OrbitCab's
#                    sound, so it is the owner's decision and not a technical edit. RT modules may only
#                    ever gain an ADDITIVE deterministic path through the SystemMath/DetMath type policy.
#   retain-local     the result never leaves the machine that computed it — a GUI curve, a pixel mapping,
#                    a display ballistic. Nobody diffs it against another row because nobody can.
#   retain-offline   offline, and the result DOES leave (a written file, a measured IR), but it is compared
#                    with a tolerance rather than byte for byte, and no parity harness reads it.
#   owner-decision   it crosses a boundary AND it is shipped bits. Both things are true, so it is listed
#                    here rather than quietly converted or quietly ignored.
#
# WHAT THIS MANIFEST DOES NOT CLAIM. It records where the calls are and what they are for. It does not
# claim to know their CADENCE: whether a call runs once per prepare() or once per sample is a property of
# the caller's loop, and no text scanner can see through a template, a helper or a call graph to decide it.
# Where cadence is the reason for a disposition it is written into the reason in words, having been read.

# --- analysis: display and live meters. None of these is in a parity harness. ---
modules/analysis/include/felitronics/analysis/CorrelationMeter.h  [exp*1]  retain-local  one setWindow() smoothing coefficient for a correlation needle; the needle is drawn, never serialised
modules/analysis/include/felitronics/analysis/MultiResSpectrumPane.h  [cos*1 exp*1 exp2*1 log10*1 log2*3 pow*1]  retain-local  spectrum analyser pane: Hann window, tilt, octave blending and dB mapping for pixels on one screen
modules/analysis/include/felitronics/analysis/MultiResSpectrumPaneFast.h  [cos*1 exp2*1 log10*1 log2*4 pow*2]  retain-local  the same pane with a precomputed plan; same pixels, same reasoning
modules/analysis/include/felitronics/analysis/PlotMap.h  [log10*5 pow*1]  retain-local  frequency<->x pixel mapping for a plot. xToFreq() does return a PARAMETER value when a user drags, but the parameter is the user's gesture, not a measurement anyone compares across rows
modules/analysis/include/felitronics/analysis/SpectrumPane.h  [cos*1 log10*1 log2*1]  retain-local  the simple pane: one Hann window, one dB conversion, one tilt — all for drawing
modules/analysis/include/felitronics/analysis/TruePeakMeter.h  [gainToDb()*1 pow*1 sin*1]  retain-rt  the LIVE meter (12 taps/phase, built for a moving needle), not the reference instrument a file is certified with. Its taps and decay ballistic are RT and shipped; ReferenceTruePeakMeter is the one in the zone
modules/analysis/include/felitronics/analysis/offline/SpectrumCurve.h  [log10*2 pow*3]  retain-offline  offline log-spaced display curve: band edges and dB. Consumed by blend/Overlay and the UI; no parity harness reads it

# --- blend / convolution: offline IR work, tolerance-compared ---
modules/blend/include/felitronics/blend/BlendKernels.h  [abs*2 cos*3 sin*2]  retain-offline  Hilbert phase rotation and crossfade kernels for IR blending; the output is an IR file, compared by tolerance in the blend suites
modules/blend/include/felitronics/blend/IrBlend.h  [pow*2]  retain-offline  dB<->linear for blend weights the user sets; the result is a written IR
modules/blend/include/felitronics/blend/Overlay.h  [pow*2]  retain-offline  overlay curve weighting for the display of a blend
modules/convolution/include/felitronics/convolution/CabConvolver.h  [log10*2 pow*2]  retain-rt  the RT cabinet convolver's IR normalisation floor and gain; this is OrbitCab's signal path
modules/convolution/include/felitronics/convolution/IrResampler.h  [sin*1]  owner-decision  the resampler kernel's sinc - one call per window MISS since P70's phase memo, not one per tap. Audited by P80 and deliberately left on the system call: these are shipped bits (OrbitCab's IR at load), pinned against a frozen unmemoized kernel in ResamplerTests, so converting it is the owner's decision and an additive one

# --- core: shared headers, mixed RT and offline consumers ---
modules/core/include/felitronics/core/DeliveryResampler.h  [pow*1 sin*1]  retain-offline  the delivery-rate resampler's Kaiser beta and its 64-tap sinc, built once per prepare(). Its output is a rendered file compared by tolerance, and P57 pinned its latency by construction rather than by these coefficients
modules/core/include/felitronics/core/Fft.h  [cos*1 sin*1]  retain-rt  the RT float FFT seam's scalar reference backend. Same twiddle-seed shape as core::offline::fftInplace, but this one is the partitioned convolver's, i.e. shipped audio. A deterministic route here would be a BasicScalarRadix2Real<Math> beside it, not a replacement — additive, and the owner's call
modules/core/include/felitronics/core/Smoother.h  [exp*1 pow*1]  retain-rt  parameter-glide coefficient (setTimeMs) and coeff^n for a fixed-grid stride. Both are audio-thread parameter smoothing in every shipped plugin
modules/core/include/felitronics/core/StreamResampler.h  [sin*1]  retain-rt  one tap of the windowed-sinc kernel, called to BUILD the kernel table in reset(), not per sample; the table then filters shipped audio

# --- dynamics: coefficients and curves on the audio thread ---
modules/dynamics/include/felitronics/dynamics/AutoLeveler.h  [exp*1 log10*1 pow*3]  retain-rt  target/threshold conversions and a per-block float log10 on the levelling path
modules/dynamics/include/felitronics/dynamics/BandBallistics.h  [pow*1 sin*1]  retain-rt  per-band attack/release and crossover design for the multiband dynamics
modules/dynamics/include/felitronics/dynamics/EnvelopeFollower.h  [exp*1]  retain-rt  the follower's attack/release coefficient — exp(-1/(tau*fs)), recomputed on setTimes()
modules/dynamics/include/felitronics/dynamics/GainReductionFollower.h  [exp*1]  retain-rt  the same coefficient for the gain-reduction follower
modules/dynamics/include/felitronics/dynamics/NoiseGate.h  [exp*1 pow*3]  retain-rt  gate thresholds (dB->linear) and ballistics coefficients
modules/dynamics/include/felitronics/dynamics/RelativeLevel.h  [exp*1 log10*1]  retain-rt  a relative-level detector's coefficient and its dB readout
modules/dynamics/include/felitronics/dynamics/TransientShaper.h  [dbToGain()*1 exp*1]  retain-rt  the shaper's envelope coefficient

# --- eq: the biggest group, and all of it is coefficient design for shipped filters ---
modules/eq/include/felitronics/eq/EqBand.h  [abs*1 polar*1 sin*1]  retain-rt  Butterworth Q design and the z^-1 phasor for a band's response. std::polar is cos+sin in disguise, which is why the lint governs it
modules/eq/include/felitronics/eq/EqEngine.h  [abs*3 log10*2]  retain-rt  the composite magnitude response in dB, for the curve the user sees and for the engine's own fitting
modules/eq/include/felitronics/eq/MatchedBiquad.h  [abs*25 cos*10 cosh*1 exp*7 log10*1 polar*1 pow*11 sin*12 sqrt*21 tan*4]  retain-rt  TabbyEQ's filter design, all of it: matched-pole placement, shelf solving, notch cascades, bandpass cascades, first-order shelves. These transcendentals ARE the coefficients, and the coefficients ARE the product's sound. The 25 std::abs are a MIX — most are scalar magnitude on a coefficient, and only the ones taking a std::complex are hypot; the 21 std::sqrt include complex square roots in the pole solver (:400, :641), which scalar IEEE exactness does not cover
modules/eq/include/felitronics/eq/MultibandSplitter.h  [pow*1]  retain-rt  the default band frequencies, laid out geometrically

# --- limiter / saturation: shipped audio, some of it per sample ---
modules/limiter/include/felitronics/limiter/TruePeakLimiter.h  [dbToGain()*3 exp*2 gainToDb()*2 log*1]  retain-rt  the release coefficient, and the ceiling glide's one-pole coefficient (exp, once per prepare()); the clip level riding a gliding ceiling reuses apply()'s two dbToGain calls through clipAt(), per oversampled sample only while a lower ceiling glides in. NOTE the limiter's other libm reach is through core::gainToDb and core::dbToGain INSIDE its oversampled per-sample loop — a carrier, not a direct call, and the reason gainToDb was not converted in place. K13 added dbToGain twice in apply(): the clip level and the knee half-width, COEFFICIENTS computed once per parameter change, which is the same consumer class as every other gain coefficient here and not a reported value. The clipper's own reported dB goes through gainToDbDet instead, because that one crosses the ABI and an ulp there moves a histogram bin
modules/saturation/include/felitronics/saturation/Saturator.h  [dbToGain()*2 exp*2 pow*2 tan*2 tanh*2]  retain-rt  drive and bias mapping for the saturator; the two model shapes' filters, designed in applyDc — at prepare and at every parameter landing, never per sample (exp for the DC blocker and the Transformer's 40 Hz flux pole, tan for Tape's two prewarped emphasis corners, pow for its shelf); and PER-SAMPLE tanh(k·L), the Transformer's core on its flux state, settled and in the glide — the same curve the WaveShaper holds for it
modules/saturation/include/felitronics/saturation/WaveShaper.h  [atan*3 tanh*14]  retain-rt  PER-SAMPLE libm on the audio path: tanh and atan ARE the waveshaping curve (Tube, Transformer and Tape are tanh curves too: Tube's through tubeRaw). Nothing deterministic can go here at any price — det has no tanh or atan, and the cost would be paid every sample. shapeAt<S>() is the same curve at explicit coefficients (atan*1 tanh*4): the Saturator's parameter glide evaluates it per sample while the coefficients move, and processSample() is untouched

# --- lineareq: FIR design, offline-built but shipped as audio ---
modules/lineareq/include/felitronics/lineareq/LinearPhaseEq.h  [cos*3]  retain-rt  the design window for a linear-phase FIR the plugin then convolves with
modules/lineareq/include/felitronics/lineareq/MagnitudeCurve.h  [cos*1 log*2 pow*2]  retain-rt  the magnitude curve a minimum-phase FIR is built from
modules/lineareq/include/felitronics/lineareq/MixedPhaseFir.h  [cos*1 log*3 sin*1]  retain-rt  mixed-phase FIR construction (cepstral, hence the logs)
modules/lineareq/include/felitronics/lineareq/NaturalPhaseEq.h  [abs*1 cos*2]  retain-rt  natural-phase design window and response

# --- measurement: the capture chain. Offline, tolerance-tested, and some of it is decisions ---
modules/measurement/include/felitronics/measurement/CaptureGate.h  [log10*1]  retain-offline  a dB reading of the captured spectrum used to decide whether a capture is usable; a decision, but one a human confirms and no parity harness reads
modules/measurement/include/felitronics/measurement/IrPost.h  [pow*1]  retain-offline  post-processing gain for a measured IR
modules/measurement/include/felitronics/measurement/LevelProbe.h  [exp*1 pow*1]  retain-offline  the probe's stimulus level and its smoothing; feeds the capture level a human sets
modules/measurement/include/felitronics/measurement/PeakClip.h  [log10*1]  retain-offline  a clip report in dB for the capture UI
modules/measurement/include/felitronics/measurement/ReferenceUnity.h  [log10*1 pow*2]  retain-offline  the -60 dB normalisation floor that decides whether a measured IR is normalised at all. Its header says plainly that this float arithmetic defines shipped audible unity gain, so it is read and left, not waved through
modules/measurement/include/felitronics/measurement/Sweep.h  [cos*2 exp*2 log*2 sin*1]  retain-offline  the exponential sweep and its inverse filter — sin() PER SAMPLE of the generated stimulus, which is the one place a deterministic sin would cost real time for a signal that is measured, not compared


# --- tools that are not on the parity surface ---
tools/eqviz/eqviz.cpp  [abs*5 arg*1 cos*1 log*8 log10*1 pow*26 sin*1 sqrt*10]  retain-local  a developer visualiser; it draws curves and nobody diffs its output
tools/fftbench.cpp  [exp*6]  retain-local  benchmark input generation


# --- carrier-only files: no direct std:: call, but they reach libm through core::gainToDb /
# --- core::dbToGain. This group is why the lint tracks carriers at all: a by-name inventory
# --- of `std::` shows nothing here, and every one of these is a shipped per-sample path.
modules/deesser/include/felitronics/deesser/DeEsser.h  [dbToGain()*1 gainToDb()*1]  retain-rt  the de-esser reads its level in dB and writes its gain back in linear, once PER SAMPLE in sharedStep(); both conversions are shipped audio
modules/dynamiceq/include/felitronics/dynamiceq/DynamicEqBand.h  [gainToDb()*1]  retain-rt  one gainToDb PER SAMPLE in sharedStep(), narrowed to float before the curve; shipped audio
modules/dynamics/include/felitronics/dynamics/Compressor.h  [dbToGain()*1]  retain-rt  dbToGain on the per-sample makeup/output path
modules/dynamics/include/felitronics/dynamics/GainReductionPath.h  [gainToDb()*2]  retain-rt  the canonical per-sample site: step() is "a log10 and a knee per sample" in its own words, and advanceSilence() evaluates the same expression once at the floor. This is the path that makes gainToDb untouchable
modules/dynamics/include/felitronics/dynamics/offline/DetectorAnalysis.h  [gainToDb()*3]  retain-offline  offline, but the histogram conversion at :169 runs per sample and the whole point of the class is to model the SHIPPED chain exactly; converting the final peak/RMS alone would make it a mixed-policy answer that is deterministic nowhere
modules/dynamics/include/felitronics/dynamics/offline/ThresholdSolver.h  [gainToDb()*1]  retain-offline  one gainToDb per frame feeding a > statisticsGateDb decision; it drives the SHIPPED GainReductionPath, so it must read what that path reads
