[LauraGPT]

Summary

Addresses the residual mixed-script case in #3796. PyPI 1.4.16 drops Latin/digit pinyin keys; current main already fixes that part, but still turns 这个端口 into 这个C端口 for fuzzy hotword C端.

Require a complete phonetic key when a Latin-containing target replaces Chinese-only input. Preserve 爱助手 -> AI助手, AI助守 -> AI助手, C段 -> C端, explicit mappings and same-script fuzzy matching. No threshold or scoring changes.

Verification

• Original 1.4.16 wheel verified against official PyPI SHA-256; the reported bug reproduces.
• New regression on unchanged main: 1 failed, 15 deselected, reproducing C-prefix insertion.
• Exact candidate: 22 passed across tests/test_postprocess_hotwords.py and tests/test_rich_transcription_postprocess.py.
• Compile of checked-out source/tests and git diff --check passed.
• Python 3.12.3, pypinyin 0.55.0, rapidfuzz 3.14.5, pytest 8.1.1; isolated Shanghai environment.

These are pure-text CPU tests, not model/GPU inference, hosted CI or full-suite acceptance. The code only changes text postprocessing. Keep #3796 open for released-version/reporter verification.

Published using the official createCommitOnBranch API with an expected-head guard and GitHub-verified signature plus DCO. The complete published tree is identical to the tested local candidate. No TLS verification, hook or repository protection was disabled. FunASR main review-count protection was explicitly changed to zero by the repository administrator on 2026-10-28; other protections remain enforced.

Current-main integration and hosted CI

• Exact head 8f2701fcf2d70320b943c596349612ba26f29bf0 has passing Python 3.10 and 3.12 text-postprocessing CI; each executes 22 tests with real pypinyin/rapidfuzz dependencies.
• Current-main 19e33badf1dba6f7606c9410c32c622486b5df13 integration tree 1e028d232f7a3e14d4942aa89d95c85d43710c29 was independently reconstructed and all 4154 blobs verified; 22 focused tests pass with no skips.
• Adds a read-only, path-filtered workflow covering hotword and rich transcription modules in all three packages. No model downloads or runtime changes.
• These are focused CPU text regressions, not pretrained ASR/GPU inference or full-suite acceptance.
• The former potential merge used pre-release main c9c42c0 and was rejected. The current potential merge parents and complete tree are verified against 1.4.17 main.
• This fix is not part of the already-published 1.4.17 packages. Keep postprocess_hotwords: non-Chinese hotword maps to an empty pinyin key and force-replaces every latin/digit span (ratio("","")=100) #3796 open for released-version/reporter verification.