Prevent glossary prompts from truncating dense opening speech

The first production Agder transcription (video edit 9590) drops the opening roll-call names and mis-times the remaining first-window text. The output becomes correct at about 00:30 with `Margit Dahle`. Evidence from an identical 75-second L4 comparison: - With the municipality glossary prompt, the first VAD window begins with `har forfalt som varamøter Reidar Saga...` and omits the preceding elected-member names. - Without the prompt, the first window correctly contains the missing names and timings through `Tor Jørgen Kjørholm`, but later misses/cases names that prompting recovers. - The prompt is 235 tokenizer tokens and does not exhaust the 448-token model sequence; the failure is prompt conditioning on the first dense VAD window, not canonical WebVTT or Kaukus chunk offsets. Intended fix: - Do not condition the first Norwegian VAD window of each media chunk on the glossary prompt; keep prompting subsequent windows and Danish behavior unchanged. - Verify a hybrid first-window/no-prompt run preserves the complete opening while retaining `Margit Dahle` and later glossary gains. - Reprocess and republish corrected chunk 0 / combined WebVTT for video edit 9590 without duplicating billable production work unnecessarily. Verified implementation result: - A naive mixed-prompt batch was rejected by the inference backend, so the worker now runs only the first prompted VAD span a second time without a prompt at batch size 1 and replaces that one raw segment before alignment; all later prompted windows and Danish behavior remain unchanged. - The identical 75-second L4 hybrid contains the complete opening roll call with correct timing through `Tor Jørgen Kjørholm` at 00:29 and retains prompted `Margit Dahle` at 00:30. - Chunk 0 was reprocessed on L4. The repaired result uses the new opening through 00:29 and the original production result byte-for-byte from the 00:30 boundary onward; the combined 4,198-cue WebVTT is republished in Kaukus. - Temporary media/routes/results were removed, the probe worker and SBS volume were deleted, production is restored to `min_workers=0`/`max_workers=1`, and only the control server's in-use root volume remains. - Worker lint, strict typing and all 12 tests pass. Murmur `41ba573` and Ansible `a97a04a` are pushed and deployed with worker digest `sha256:31a07d2dab05aa37ba1dc878d3b1192a07e948f8fcaf49ca7b2bf1fdce6b2707`; health is 204 and the final Ansible rerun had zero changes. Acceptance criteria: - Regression tests cover per-request prompt scoping and first-window prompt suppression for nb/nn only. - Identical first-75-second L4 output contains the missing opening names at correct timings and retains prompted output from 00:30 onward. - Full affected transcript is rebuilt and inspected around the repair boundary. - Temporary media, routes and diagnostic files are removed.