fix: place pauses at every sentence, repair modern weight norm checkpoints

[thewh1teagle]

Three bugs found while synthesizing a Hebrew book with continuous=True, all of which failed silently.

Newlines were deleted, not turned into gaps

\n is not in the vocabulary, so it was dropped with nothing in its place and consecutive lines fused:

mimˈul?""kˈen.""ʔˈat        what the model received
mimˈul?" "kˈen." "ʔˈat      after the fix

With no word boundary there was nothing for the model to pause on, so a line of dialogue ran straight into the next. Whitespace is now collapsed to single spaces in _prepare.

Only continuous synthesis was affected: the chunker splits on \s+ and rejoins with spaces, so batched mode never saw it.

Pauses existed only at batch joins

sentence_pause and clause_pause were applied where two batches met and nowhere else, so a sentence inside a batch got only what the model produces on its own, about 0.1s, and continuous synthesis got nothing at all.

The new pauses module places them with the timings instead, after every mark wherever it falls. It tops up rather than inserts blindly: a mark is often timed slightly before the gap it causes, so it looks for the pause already there in a short window past the mark and adds only the difference.

Measured on the Hebrew book, 33 sentence marks over 79s:

	before	after
gap after a line of dialogue	0.12s	0.31s
median sentence pause	n/a	0.48s
with sentence_pause=0.45		0.51-0.82s

Models with no duration output keep the previous batch join behaviour, since there are no timings to place anything with.

Checkpoints trained on modern torch exported to static

KModel registers weight norm as weight_g/weight_v, while checkpoints saved by recent torch store parametrizations.weight.original0/1:

model expects:   decode.0.conv1.weight_g, decode.0.conv1.weight_v
v1.0 ckpt has:   module.decode.0.conv1.weight_g, ...weight_v                    375/375 load
modern ckpt has: module.decode.0.conv1.parametrizations.weight.original0/1      235/375 load

KModel catches the failure and retries with strict=False, so the other 140 decoder tensors stayed randomly initialized and the exported model produced static, with no error at any point. verify would not have caught it either: it compares ONNX against torch, and both were noise.

export.py now rewrites those keys before loading, and is a no-op for checkpoints that do not need it. Confirmed by re-exporting the Hebrew checkpoint: spectral flatness 0.712 (noise) before, 0.109 after, and the result is indistinguishable from a known good ONNX of the same model (15.10s both, rms 0.1318 vs 0.1317).

Verification

Full suite passes, save.py and with_stream_save.py still run.