fix: align repository SenseVoice timestamps with FunASR

[LauraGPT]

Summary

Related to #291. This does not close the issue: the reporter's original audio and pinned environment are unavailable.

• Initialize the repository position encoder as an nn.Module so checkpoint state inspection and normal module calls work.
• Bring the repository's timestamp output into FunASR's VAD contract: millisecond pairs plus a parallel words list, instead of token/seconds triples that fail at int(token).
• Supply CTC log probabilities and re-encoded text targets; retain original-token ownership when one token expands into multiple IDs, and reject incomplete alignment instead of silently truncating it.
• Keep batch lengths per item and normalize SentencePiece word boundaries.
• Document the direct-call schema migration and update demo2.py; add CPU model contracts to CI.

Evidence

All development and real inference ran on ind-gpu8. Model weights were not changed.

• Original repository implementation: actual checkpoint loading fails with missing nn.Module state hooks under PyTorch 2.11.0+cu128.
• Module-only correction: both punctuation-on and punctuation-off pipelines fail with invalid literal for int() on a punctuation token, matching the timestamp type failure shown in the reporter's second screenshot.
• A schema-only intermediate candidate was rejected: it returned only 3 aligned words for a full transcript. That evidence is retained; eliminating the exception alone was not acceptance.
• A multi-ID real-aligner regression failed before the owner map, producing [[0,30],[30,90]] instead of [[0,90],[150,210]].
• Final candidate: 44 repository tests passed with third-party pytest plugin autoload disabled; compile and whitespace checks passed.
• Actual published FunASR 1.4.15 wheel was isolated and all 443 RECORD hashes checked. Fixed SenseVoice HF revision3847d57b6bdf2dd8875cb1508d2af43d80a16bf7, FSMN-VAD df20e6b, CAM++ e4b6ede7 and CT-punc d0e55e2 were used.
• On the fixed public 51.66275-second sample, final repository source loaded successfully and returned 243 aligned words/timestamp pairs covering all decoded text with punctuation both off and on. The composed pipelines produced 3 and 25 sentence records respectively.
• Actual unequal-length no-VAD batch: English7.18s and Chinese18s returned15/76 word pairs with complete text coverage and valid chronological intervals. Timestamp-disabled output retained its original schema.

Limits

This is functional/schema validation, not word-boundary accuracy, diarization accuracy, speaker identity recognition, hardware throughput, or a fix for the reporter's first empty-result screenshot. The model still emits Korean text on a one-second zero waveform without VAD; that existing hallucination is not hidden or claimed fixed. Punctuation duplication in the public sample is also not fixed here.

The tested runtime is a preserved PyTorch2.11/FunASR1.4.15 environment, not a clean installation of this repository's newer default requirements. CPU CI separately tests module/CTC contracts on PyTorch2.12.1; the normal container build validates installation. Original model card/Space permissions and GHCR visibility are unchanged.

Rollback

The original branch bundle and every candidate/source/log are backed up under the operations evidence store. Only the six explicitly reviewed files are included. No unrelated checkout or dependency changes, model publication, PyPI release or website deployment.