Feature: Add Filtering by Log-Probability (Whisper Confidence)

All transcripts in the system come with token-level log-probabilities from the Whisper model. These values reflect the model’s confidence and can help users surface only high-certainty matches.

**Proposal**:
Add an optional filter that lets users restrict search results based on a minimum log-probability threshold (e.g., filter out tokens below –3.0). This can be exposed as:
- a numeric slider (log-prob range), or
- preset buttons (e.g., all, medium confidence, high confidence).

**Why it’s useful:**
- Reduces noise from low-confidence ASR segments.
- Enables more trustable keyword/regex searches.
- Helps with downstream workflows where accuracy of the match is important.

**Implementation notes:**
- Filtering should occur after match extraction, using per-token or segment-level log-probs already stored in the DB.
- Consider exposing average log-prob per snippet, and allow filtering by snippet-mean vs. token-min.
- Ensure UI clearly indicates when confidence filtering is active.

Start by tracing search match extraction and the stored per-token or segment-level log-probabilities from the Whisper transcripts. Resolve whether filtering uses snippet means or token minima, then identify the UI entry point for a slider or presets. Done means results can be filtered by the chosen threshold and the UI clearly indicates when confidence filtering is active.

Written by the indexing model from the issue text.