fix(stt): wire stt.local.initial_prompt through to faster-whisper

[Akloenx123]

Summary

faster-whisper exposes an initial_prompt kwarg that biases the model's
vocabulary toward domain-specific terms — useful for transcribing voice notes
that contain product names, technical jargon, or other words the default model
rarely sees in training (e.g. "DeepSeek", "MiMo", "Kicktipp", "Cronjobs").

Before this change, only stt.local.language was wired through from
config.yaml to the transcribe() call. initial_prompt was silently
ignored even when set, forcing users to either patch the code or accept
garbled transcripts of technical vocabulary.

This commit reads stt.local.initial_prompt from config.yaml and passes
it as the initial_prompt kwarg to faster-whisper's transcribe(). The
setting is optional and defaults to absent — existing configs are unaffected.

Reproduction (before)

# config.yaml
stt:
  provider: local
  local:
    model: large-v3
    language: de
    initial_prompt: "DeepSeek, MiMo, MiniMax, Kicktipp, Cronjobs"

User sends a voice note saying "schick mir den DeepSeek V4 Pro Tipp".
Transcribed result: "schick mir den Debsig V4 Pro Tipp".
The initial_prompt line in config is silently ignored.

After this fix, the same voice note transcribes correctly: "schick mir den
DeepSeek V4 Pro Tipp".

Why this should merge

1. Minimal blast radius — 8 lines of core code, all additive.
2. Backwards compatible — initial_prompt defaults to absent; existing
   configs behave identically.
3. Solves a real reported problem — voice-to-text for technical vocab
   (model names, product names, internal jargon) is a recurring pain point.
4. Mirrors existing language plumbing — same config-yml pattern,
   same read order, same opt-in shape.
5. Tests included — three contracts: prompt set / unset / empty string.

Test plan

• tests/tools/test_stt_initial_prompt.py — 3 tests, all passing
• Existing tests/tools/test_transcription.py — unchanged behavior
• Manual E2E with large-v3 and a domain-specific voice note:
  • Before: "Debsig", "Kick-Tip", "Conjobs"
  • After: "DeepSeek", "Kicktipp", "Cronjobs"

AGENTS.md alignment

This change follows the project's "narrow waist" guidance:

• No new core tool. Uses existing stt.local config namespace.
• No new HERMES_* env var. All wiring is config.yaml-only, per the
  "reject PRs that tell users to set X in your .env" rule.
• Extend, don't duplicate. Reuses the existing _load_stt_config() /
  _forced_lang / transcribe_kwargs plumbing — no parallel code path.
• Behavior contract tests, not snapshots. Tests assert the kwarg-dict
  invariant (initial_prompt is present iff config sets it), not frozen
  model output.

Happy to address review feedback or split this into smaller commits if
preferred.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #50684 — same mechanism: both wire stt.local.initial_prompt from config.yaml through to faster-whisper's transcribe() in tools/transcription_tools.py to bias vocabulary toward domain-specific terms, both targeting #50181. #50684 is the earlier open PR. Cross-linking so a maintainer can pick the canonical one.

[Akloenx123]

Closing in favor of #50684 — opened earlier (2026-06-22 vs 2026-06-28), smaller diff (+50 vs +118), includes the cli-config.yaml.example doc update I missed, and handles the whitespace edge case (initial_prompt.strip()) my patch skipped.

Thanks to @BlackishGreen33 for getting there first. Marking as superseded so a maintainer only has to review one PR.