Blocked on #1 — implement in the frequency order d2rmv curves reports, not all fifteen on spec.
Six of Granny's 21 curve encodings are decoded today (DaIdentity, the three constant forms, DaKeyframes32f, DaK32fC32f). The rest quantise knots and controls into 8- and 16-bit fields. They are recognised and reported, and GrannyCurve.Decode throws GrannyCurveException naming the format rather than guessing.
That refusal is deliberate: a wrong quantisation layout does not raise an error, it produces animation that plays at the wrong speed or twists the wrong way — discovered in the SC2 editor weeks later, not in a stack trace.
Three things to confirm against real data first
Detail in docs/GR2-FORMAT.md.
- Knot reconstruction. These formats store
OneOverKnotScaleTrunc as a uint16 holding the high half of a float. Whether the reconstructed value multiplies or divides the raw knots decides whether a clip plays at the right speed — and the field's name and its usage disagree across public sources.
D4nK16uC15u bit packing. A "smallest three" quaternion: three 15-bit components plus a selector for the dropped one, packed into 48 bits. Which bits hold the selector, and how the dropped component's sign is recovered, needs reading off real data.- Degree. A degree-2 curve is a quadratic B-spline with more controls than knots; sampling at its knots is not evaluating it.
GrannyCurve.Decode currently refuses a DaK32fC32f whose control count is not a whole multiple of its knot count, reporting both numbers. If real files hit that, spline evaluation is the work.
Route
Norbyte/lslib is MIT and has working curve decoders for Divinity and Baldur's Gate GR2s. Adapting the ones the survey names, with attribution in the README alongside the existing credits, is the intended path.
Do not copy from neptuwunium/Knit — it is EUPL-1.2, incompatible with keeping this MIT. Read it for a second opinion only.
Done when
- Each newly implemented encoding has a fixture in
SyntheticGrannyAnimation covering it.
GrannyCurve.Implemented is updated so d2rmv curves reports the new coverage.- An animated monster exported to
.glb plays in Blender at the same speed and direction as in Noesis. That end-to-end check is the one that catches a wrong scale convention; the unit tests cannot.
Blocked on #1 — implement in the frequency order
d2rmv curvesreports, not all fifteen on spec.Six of Granny's 21 curve encodings are decoded today (
DaIdentity, the three constant forms,DaKeyframes32f,DaK32fC32f). The rest quantise knots and controls into 8- and 16-bit fields. They are recognised and reported, andGrannyCurve.DecodethrowsGrannyCurveExceptionnaming the format rather than guessing.That refusal is deliberate: a wrong quantisation layout does not raise an error, it produces animation that plays at the wrong speed or twists the wrong way — discovered in the SC2 editor weeks later, not in a stack trace.
Three things to confirm against real data first
Detail in
docs/GR2-FORMAT.md.OneOverKnotScaleTruncas auint16holding the high half of a float. Whether the reconstructed value multiplies or divides the raw knots decides whether a clip plays at the right speed — and the field's name and its usage disagree across public sources.D4nK16uC15ubit packing. A "smallest three" quaternion: three 15-bit components plus a selector for the dropped one, packed into 48 bits. Which bits hold the selector, and how the dropped component's sign is recovered, needs reading off real data.GrannyCurve.Decodecurrently refuses aDaK32fC32fwhose control count is not a whole multiple of its knot count, reporting both numbers. If real files hit that, spline evaluation is the work.Route
Norbyte/lslibis MIT and has working curve decoders for Divinity and Baldur's Gate GR2s. Adapting the ones the survey names, with attribution in the README alongside the existing credits, is the intended path.Do not copy from
neptuwunium/Knit— it is EUPL-1.2, incompatible with keeping this MIT. Read it for a second opinion only.Done when
SyntheticGrannyAnimationcovering it.GrannyCurve.Implementedis updated sod2rmv curvesreports the new coverage..glbplays in Blender at the same speed and direction as in Noesis. That end-to-end check is the one that catches a wrong scale convention; the unit tests cannot.