quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema.
Declare the pipeline as one multi-component variant
The duration predictor and the synthesizer run together; they were declared as
sibling variants, which forced both to be default: true (xnnpack) or picked one
arbitrarily (coreml). They are now one variant with components, their methods
nested per component. Dynamic dimensions use -1, as the schema specifies and
every other repository already does.
Add the Core ML models, and repair .gitattributes
main and the v0.10.0 tag had diverged: the Core ML duration predictors and
synthesizers (standard, polish, german) were only ever pushed to the tag, so
main did not describe what ships. This brings main up to the tag.
.gitattributes was also broken on the tag side. main's copy had no trailing
newline, so appending to it fused two rules into one line:
voices/** filter=lfs diff=lfs merge=lfs -textcoreml/standard/duration_...
That left `voices/**` with a malformed attribute and dropped the `finetunes/**`
rule entirely. Both are restored here and the six Core ML rules are on their
own lines.
The .pte contents are unchanged: these are the same LFS objects the tag
already references.
