audiodynamics DYN_TRANSIENT: the +1e-5f envelope floor makes the differential level-dependent in a note's tail

[bdbarnett]

Found by the TransientShaper gate-audit fix round (audiocomponents,
effects/p2b-transientshaper, 2026-09-07). This is a finding, not a
regression — the code is as it has always been, and the class ships on it.
It is filed because the class's evidence pack names it as the measured cause
of a trait it had to disconfirm, and no issue existed.

Stakes. DYN_TRANSIENT's whole promise is that it is thresholdless:
both control voltages are differences between two envelopes of the same
signal, so the level cancels and a ghost note is shaped as hard as a
rimshot. That is exactly what TransientShaper's docstring sells. It stops
being true near the floor.

The mechanism. src/shared/audioif_dynamics.c:628-630 takes
gain_to_db(env + 1e-5f) on each envelope. 1e-5 is −100 dBFS, so it is a
pedestal added before the log, and the difference of two logs is scale-free
only while both envelopes are far above it. It is 1 % of a −60 dBFS envelope
and 10 % of a −80 dBFS one.

Measured, on audiocomponents/.venv/bin/python at the pin 2f6cbc3,
48 kHz, through audioeffects.rebuilt.transientshaper:

• peak attack gain at Attack +12, one hit at four levels:
  +12.000 / +12.000 / +11.996 / +11.953 dB at −6 / −20 / −40 / −60 dBFS,
  and +11.076 dB at −80 dBFS — a graded departure that tracks the
  pedestal's share of the envelope, not anything in the gain law.

• it is not only a −60 dBFS effect. The tail of a −40 dBFS note gets down
  near the pedestal too, so the point-by-point spread between the same hit at
  −20 and −40 dBFS depends on how deep into the decay it is read. Same
  renders, read to N dB under each probe's own peak
  (tools/phase2_probes/transientshaper_patches.py):

  Read depth	patch 0 Flat (a near-wire)	patch 2 Room Off	patch 5 Kick Punch	patch 6 Ambient Swell
  10 dB	0.043	0.032	0.042	0.224
  20 dB	0.089	0.101	0.094	0.441
  30 dB	0.134	0.630	0.472	0.803
  40 dB	0.142	2.797	1.578	1.322
  50 dB	0.142	6.588	3.068	1.322

  The bar the class's trait sets is 0.500 dB. The near-wire settles and stops;
  the three working patches do not.

What the class did about it: nothing to the DSP. T1 is recorded
disconfirmed as stated and the surviving claim is bounded in the docstring
and the catalogue row to where it holds. The full record is
audiocomponents/docs/effects/TransientShaper-evidence.md §1 and §11.

The ask, if any. A smaller pedestal, or a floor applied to the ratio
rather than to each envelope, would push the departure below any level a
16-bit signal reaches. NEED: NONE — nothing in audiocomponents is blocked on
this; the class ships with the claim bounded. Filed so the number exists on
this side of the fence.

[bdbarnett]

I built both remedies this asks for, measured them, and both are worse. Nothing was added to the node; 87bfdde records the measurement beside the constant so it does not get repeated.

What I measured

The property DYN_TRANSIENT promises is that the control voltage depends on the shape, not the level — both envelopes are of the same signal, so the level cancels. So: render one shape at a reference level, render it again much quieter, and take the worst departure between the two gain traces. Swept over the pedestal, at −30 / −40 / −50 / −60 dBFS against the same shape at −6:

  pedestal          -30      -40      -50      -60
  1e-3           0.3690   1.1117   4.4914   7.4214
  1e-4           0.0361   0.1185   0.3671   1.0456
  1e-5           0.0022   0.0066   0.0208   0.0448   <- what ships
  1e-6           0.0043   0.0133   0.0387   0.1000
  1e-7           0.0045   0.0140   0.0410   0.1093
  0 (exact ratio) 0.0045  0.0141   0.0412   0.1103

Non-monotonic, with the minimum at the shipped 1e-5. Both asks land on the wrong side of it: "a smaller pedestal" is worse from 1e-6 down, and "a floor applied to the ratio rather than to each envelope" — which I implemented as one log of fast/slow, exactly scale-free in arithmetic — is the worst row but one.

Why, and why the issue's reasoning is half right

The mechanism this issue names is real: a pedestal added before the log only cancels while both envelopes are well above it, so it makes the differential level-dependent. That is why 1e-3 and 1e-4 are bad.

What it misses is that the pedestal is also damping int16 quantisation noise, which a bare ratio amplifies. Adding the same constant to numerator and denominator pulls the ratio toward 1 and desensitises it to small relative errors in two envelopes of an already-quantised signal. The two effects pull opposite ways, and 1e-5 is where they cross.

What I could not reproduce, which matters for the report

I could not reproduce this issue's headline number at the node. A bare Dynamics(DYN_TRANSIENT, attack_gain_db=12) reads +12.000 dB at every level from −6 to −80 dBFS, where the issue reports +11.076 at −80 — because a hit sharp enough to be a transient drives the differential past the ±6 dB window the gain law normalises over, so it clamps and the pedestal cannot show. The issue's figures were taken through audioeffects.rebuilt.transientshaper, and it may be the class's chain — its own envelope, its patch settings — that puts the differential inside the unclamped window. That is worth establishing before anyone acts on the numbers, and it is the kind of thing that would change an upstream report, which is why it is being said here rather than filed anywhere.

Where that leaves it

The issue's own standing is unchanged: NEED: NONE, the class ships with its claim bounded, and the record is in TransientShaper-evidence.md. What is new is that the two proposed remedies are measured and rejected, and the reason 1e-5 is the number now lives in the source rather than looking arbitrary.

Leaving this open for you: it is either closeable as answered — the ask was investigated and declined on measurement — or it wants a proper reproduction through the class first. I would not spend more on it without that reproduction, because without it there is no defect at the node to aim at.

[bdbarnett]

Closed as investigated and declined (Brad, 2026-09-09).

Both remedies this asks for were built and measured, and both are worse: the departure is non-monotonic in the pedestal with its minimum at the shipped 1e-5, because the pedestal also damps the int16 quantisation noise a bare ratio amplifies. The full sweep is in the comment above and, more usefully, in the source beside the constant (87bfdde), so the value no longer looks arbitrary and nobody repeats the experiment.

The finding stands and the record stays where it belongs — TransientShaper-evidence.md §1 and §11 for the class's disconfirmed trait, and the code comment for why the number is what it is. What remains unestablished is whether the reported +11.076 dB at −80 dBFS reproduces at the node at all; it does not in a bare Dynamics, which reads +12.000 dB at every level from −6 to −80 because a real transient clamps the differential past the ±6 dB window. If that reproduction is ever wanted it is a fresh question with a fresh number, not this issue.