feat(voice): voice core — honest layer reach, spoken summary with on-screen card, narration, point-and-ask

- Voice layer manifest (src/voice/layerManifest.js) generates the layer enums,
  aliases and analyst field hints; a test fails when a catalog layer has no entry.
- analyst_query validates fields, operators, units, values and scopes, counts up
  to 250,000 loaded records per layer, reports complete/partial/unanswered, tells
  LAYER_OFF / NOT_READY / FEED_UNAVAILABLE apart from zero, refuses mismatched
  follow-ups and never lets a superseded query replace follow-up memory. Seven
  more layers expose analyst records.
- Tool results carry a code-built say line and a display card; the voice card in
  the mic control shows captions, plan steps and the result
  (src/voice/resultDisplay.js is the one adapter). Slow tools narrate their
  current step out of band; holding Space while the assistant talks cuts it off.
  Turn spans go to the debug log; caption transcription is metered; the log
  omits what was said unless GEV_VOICE_LOG_CONTENT=1.
- Point and ask: the Space-down pointer snapshot resolves "this", "that one" and
  "around here" through 'pointer' targets, and "the second one" through
  referent:n from the last numbered list.
- Prompt states each policy once (instructions + tools 48.5k -> 46.7k chars).
- Voice benchmark harness (scripts/qa-voice-bench.mjs, scripts/voice-bench/)
  shares its phrase table and grading with scripts/qa-voice-routing.mjs.

CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+- The voice card shows analyst counts ("At least 250,000 aircraft …") with
+  their scope, feed window and caveats, and marks partial answers.
+- Voice answers are shorter and fill slow work instead of going silent. Slow
+  or multi-step requests may open with one short spoken plan; instant ones
+  just happen. Tool results carry a code-built `say` line and a `display`
+  card: sources and notes stay on screen, a caveat is spoken only when it
+  changes the answer, and "at least" and partial answers are always said. A card above the mic shows captions, the plan's steps and the
+  result. When a tool that reports steps (layers, place lookups, outlines,
+  nearest-aircraft) is still running after 1.5 s its current step is spoken
+  once; others say "still working" once after 8 s. A soft
+  tick plays in push-to-talk. Holding Space while the assistant talks cuts it
+  off. Each turn writes latency and speech figures to the debug log
+  (`node scripts/voice-turn-spans.mjs`). Captions of what you say are
+  counted in the voice cost meter; the debug log leaves out what was said
+  unless `GEV_VOICE_LOG_CONTENT=1`.
+- Point and ask: hold Space with the cursor on an aircraft, ship or spot on
+  the map and say "what's that", "track that one" or "how many flights
+  around here". A reticle marks the point and the voice card shows what
+  "this" meant. In open mic it works when you have just moved the cursor
+  there. "The second one" picks from the last numbered list. Asking about
+  bare ground up close gives the assistant a small crop of the map around the
+  point.
+
 - Public Overpass instances are no longer used by default. Street Traffic
   roads come from TomTom flow tiles, OpenFreeMap vector tiles, or both, chosen
   on the layer row (TomTom / OSM / Hybrid) or with `?trafficRoads=`. With a
@@ -149,6 +172,7 @@
   geocode and admin-boundary fallback answers `region-timeout` after 3 s
   instead of holding the reply.
 
+
 - Transit and Directions rows repaint as soon as their data lands again:
   `refreshLayerStats()` now lives on the layer lifecycle, not only on the
   compatibility facade. `scripts/qa-radio.mjs` uses it instead of a private
@@ -182,6 +206,30 @@
   (`src/ui/imagerySplit.js`, `src/maps/imageryComparison.js`), and
   `MapSourceController.subscribe()` reports every settled map switch.
 
+## Unreleased — voice reaches every layer; honest analyst answers
+
+- Voice can switch on every shipped data layer (the two scene-driven
+  Bhote Koshi layers stay scene-only), including transit, Recent
+  Imagery, wind, the weather layers, cyclones and mapped installations, and
+  understands their spoken names ("buses", "hurricanes", "military bases").
+  One voice manifest (`src/voice/layerManifest.js`) generates the tool enums
+  and aliases, and a unit test fails when a new layer has no entry.
+- `analyst_query` counts up to 250,000 loaded records per layer instead of the
+  first 2,000, and says when a layer went past that (`complete: false`). It
+  also answers over bikeshare stations, transit vehicles, rocket launches,
+  cyclones, ALPR cameras, mapped installations, local ADS-B aircraft and fire perimeters; fire
+  records carry their sensor (VIIRS or MODIS).
+- Unknown fields, operators, units, value types, scope kinds and invalid
+  centres are refused with the allowed values instead of answering zero or widening to anywhere. A
+  layer that is off, still loading or unavailable is reported as such, not
+  as a count of zero; a partly answerable query is marked `partial`.
+  Distance sorts honour `sortDir`, a written unit such as `altitudeM(ft)`
+  converts, and large rankings no longer sort the whole set.
+- Results carry `lat`/`lon` for each item and a `display` object with the
+  scope, the feed window (earthquakes and fires cover the last 24 h) and any
+  caveat for the screen. The model is asked to answer in one short phrase.
+- The weather and Recent Imagery panels can be opened by voice.
+
 ## Unreleased — local receiver feeds
 
 - The Local ADS-B layer also reads local 1090 MHz and 978 MHz UAT decoder