[nexxtmobile]

Fixes #12748

ASR backends like Nemotron-3.5 append end-of-utterance language tags to
transcripts ("Hallo. <en-US>"). The committed transcript feeds the LLM
conversation verbatim (user item Transcript field), and the per-utterance
tag is unreliable for short/ambiguous utterances — the LLM treats it as a
language signal and switches the reply language against the session
instructions (observed: German "Hallo." tagged <en-US> → English reply in
a German call).

commitUtteranceWithTranscript now strips the EOU tags from the transcript
that goes to appendUserItem/generateResponse (the LLM context). The
emitted input_audio_transcription event keeps the raw transcript
(debugging / client display).

Includes regression specs: stripEOULanguageTags unit specs + a client-commit
spec asserting the emitted event keeps the raw tag while the LLM user message
is tag-free.

Tested: go test -race ./core/http/endpoints/openai/ green (Go 1.26).

[localai-org-maint-bot]

@mudler Code-review sign-off at 75b8bf975, subject to CI and human DCO certification of the documentation fix. The client receives the raw transcription before language-tag removal; the LLM receives the cleaned transcript, with the existing empty-transcript guard retained. I added the required documentation. Reviewed against master 48f358dd9; whitespace checks pass. Focused endpoint tests could not start because the dependency proxy denied LocalAGI; race testing also lacks CGO/a C compiler. Nonblocking caveat: the broad tag pattern can also remove literal short HTML tags such as <div> from transcription text.

[nexxtmobile]

Update: rebased onto mudler:master e51747055 (the branch had gone stale / conflicting). go test -race ./core/http/endpoints/openai/ is green on the new base.

Note on the doc commit: a rebase force-push clobbered the maint-bot's doc commit (75b8bf975). I have re-applied the identical doc change (same file, same text) as 005a2d4ce with human DCO attestation, per the process used in #12622. Sorry for the churn.