Improve callsign matching for spelled-out vs. numeric variants

[MuddyWinds]

Whisper transcribes callsigns inconsistently depending on how the controller/pilot speaks and how the audio comes through — e.g. "United two five", "United 25", and "UAL25" can all refer to the same aircraft. This fragments analysis and weakens ADS-B correlation (which matches on callsign).

Goal

Add a normalization pass that canonicalizes a callsign into a consistent form (airline telephony + flight number) before the transcript batch is sent to Gemini and before ADS-B correlation.

Suggested approach

• Map airline telephony names → ICAO prefixes (e.g. United → UAL, Speedbird → BAW, Cathay → CPA).
• Convert spelled-out and word-number digits to numeric (two five → 25, niner → 9).
• Produce a canonical token (e.g. UAL25) while preserving the raw transcript text for display.
• Be conservative: when confidence is low or no airline match is found, leave the raw text untouched rather than guessing.

Where to look

• backend/core/batcher.py (batch assembly + AIRPORT_GEO / ADS-B correlation)
• The Gemini prompt assembly path

Acceptance

• A short unit test covering several spelled-out / numeric / ICAO variants resolving to the same canonical callsign.
• Raw transcript text remains visible on the observation card.

This maps directly to the "callsign matching isn't perfect yet" limitation in the README — one of the most impactful accuracy fixes available.