fix(voice): keep canonical analyst IDs out of display-text normalization

compactAnalystItem normalized every string field, including `id`.
Bikeshare uses the raw station name as its ID, so a station with
repeated spaces or a name over 180 code units was registered as a
referent under a different key and a numbered fly_to_location reported
it as no longer loaded. Normalize display fields only; layerKey, id,
icao24, mmsi, noradId, registration and callsign pass through as the
layer reports them.

Adds an analyst_query -> referent -> fly_to_location test for
repeated-space and long station names, and documents the rule in
docs/CURRENT-STATE.md.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs/CURRENT-STATE.md

@@ -13,8 +13,12 @@ External ALPR tags, surveillance-tool display fields, voice context text and
 compact analyst strings collapse controls and whitespace before a final
 180-character cap (scene labels retain their 120-character cap). Empty ALPR
 values remain null and Unicode text is preserved without splitting a surrogate
-pair. These fields remain untrusted source data in structured tool output; text
-normalization does not establish trust or eliminate prompt injection.
+pair. Only display fields are normalized: canonical analyst identifiers
+(`id`, `layerKey`, `icao24`, `mmsi`, `noradId`, `registration`, `callsign`)
+pass through unchanged, because bikeshare uses the raw station name as its ID
+and numbered voice references resolve by that key. These fields remain
+untrusted source data in structured tool output; text normalization does not
+establish trust or eliminate prompt injection.
 
 Terrain height requests accept at most 64 points, matching the browser service's
 existing batch size and one bounded upstream chunk. Larger requests return HTTP