merge main into fix/cctv-tool-routing

docs/CURRENT-STATE.md

@@ -618,7 +618,9 @@ Reference feed construction is exported through `sources/reference`; the cable s
 
 Source factories have dedicated `layers/<family>/source` exports. ALPR and earthquake record normalization and CCTV source policy no longer pull rendering into source consumers. Catalog construction and voice feed reads use their focused owners; compatibility entries remain available. Settings filesystem hardening lives under `server/standalone/`. Import-direction checks complement export ownership checks; source behavior, settings policy and rendering are unchanged.
 
-Voice controls bind to a protocol-independent session factory. The default WebRTC adapter preserves the existing Realtime connection, push-to-talk, cost controls and radio handoff. Session subscriptions expose state, transcript, action call/progress/result, interruption, completion and turn-metrics events; stopping or replacing a session cancels pending actions. Alternate adapters can use the same controls and action runner.
+Voice controls bind to a protocol-independent session factory. The default WebRTC adapter preserves the existing Realtime connection, push-to-talk, cost controls and radio handoff. Session subscriptions expose state, transcript, action call/progress/result, interruption, completion and turn-metrics events; stopping or replacing a session cancels pending actions. An active session's normal Stop publishes one idle transition, including adapters that emit their own terminal state; a restart triggered by that transition keeps ownership of its new startup state. Alternate adapters can use the same controls and action runner.
+
+The standalone app offers OpenAI and Gemini Live through an explicit provider chooser. Gemini keeps its permanent API key on the local server and mints a short-lived, single-use credential with server-owned instructions and the standalone action catalog. Its native WebSocket adapter supports open microphone and push-to-talk input, streaming audio, captions and the existing action runner. In push-to-talk, a tool call that arrives with no input, reply or tool result outstanding is answered with a failure and never runs, and neither does any other call in the reply it starts; open-microphone turns start on Gemini's server-side activity detection, which reports no boundary to check. Switching providers ends the old conversation and releases its microphone and connection; only the input mode carries over. Gemini uses bounded shutdown draining and explicit manual restart, without automatic reconnection or a promise of uninterrupted session duration. Hosts that do not opt into the provider facade retain their existing default or custom adapter.
 
 Voice action argument schemas have one portable owner under `src/voice/actionSchemas.js`. The Realtime provider builds 30 tools using separate description-only metadata. Description customization cannot replace argument types, enum values or required fields.
 
@@ -3453,7 +3455,7 @@ its criteria cannot be silently ignored.
 | Satellites             | CelesTrak                                                                                                                                                                                       | `src/data/satellites.js`                              | `/api/celestrak`                                         | 120s                                                                              |
 | Space Missions (30d)   | Launch Library 2 + CelesTrak                                                                                                                                                                    | `src/data/rocketLaunches.js`                          | `/api/launches` + `/api/celestrak/active`                | 5 min                                                                             |
 | Traffic | Selectable TomTom / OpenStreetMap / Hybrid roads; optional TomTom flow (BYOK) | `src/data/traffic.js` | browser-direct OpenFreeMap tiles; `/api/tomtom` for flow | viewport-driven; capped tile cache |
-| CCTV                   | Austin + Caltrans (CA) + TfL London + Ontario 511 + Fintraffic (FI) + DriveBC (BC) + TxDOT (TX) + Estonia (Tallinn, Tarktee) + Live Traffic NSW + Open Calgary Open Data + Statens vegvesen (NO, stills + live HLS) + King County (WA) Open Data + WSDOT (WA) Open Data + NYC DOT TMC (opt-in) + City of Toronto RESCU + Catalonia (Servei Català de Trànsit) + FL511 (FL, opt-in) + Street View fallback | `src/data/cctv.js`                                    | `/api/cctv`                                              | 10s (active)                                                                      |
+| CCTV                   | Austin + Caltrans (CA) + TfL London + Ontario 511 + Fintraffic (FI) + DriveBC (BC) + TxDOT (TX) + Estonia (Tallinn, Tarktee) + Live Traffic NSW + Open Calgary Open Data + Statens vegvesen (NO, stills + live HLS, Kartverket ground heights) + King County (WA) Open Data + WSDOT (WA) Open Data + NYC DOT TMC (opt-in) + City of Toronto RESCU + Catalonia (Servei Català de Trànsit) + FL511 (FL, opt-in) + Street View fallback | `src/data/cctv.js`                                    | `/api/cctv`                                              | 10s (active)                                                                      |
 | Radio                  | Radio Browser (public-domain station directory)                                                                                                                                                 | `src/data/radio.js`                                   | `/api/radio/stations`, `/api/radio/click/:uuid`          | 45 min directory refresh                                                          |
 | Transit 🚌             | Operator GTFS-Realtime VehiclePositions (7 keyless regions, `src/data/transitFeeds.js`)                                                                                                         | `src/layers/transit/` via `src/app/layers/transit.js` | `/api/transit`                                           | 15s (poll + delayed playback)                                                     |
 | Bikeshare 🚲           | GBFS (Lyft + BCycle + Social Bicycles + PBSC incl. Bike Share Toronto, Bike Itaú (Brazil), Bicing (Barcelona) and seven other Spanish cities + Urban Sharing, Entur, JCDecaux (Lillestrøm, Valencia, Seville) + nextbike (AMBici))                                                                                                                                        | `src/data/bikeshare.js`                               | `/api/gbfs`                                              | 60s                                                                               |
@@ -4424,7 +4426,7 @@ silently demoting every later lookup for the session.
 - **FIRMS**: no ground clamping (zero 3D-tiles height sampling), ≤18 screen-decluttered ambient labels, click-to-inspect detail card, 2.5k/3k sprite budgets viewport-clipped by FRP.
 - **CCTV v2 foundation:** a pitched
   frustum wireframe (4 corner rays + far-cap rectangle) with a monitor plane at the frustum's
-  far cap, retargeting the existing video/canvas texture pipeline. Manual calibration only —
+  far cap, retargeting the existing video/canvas texture pipeline. Stills are drawn at their source aspect, centered on the placeholder dark (`#01070c`), instead of being stretched to the 16:9 canvas. Manual calibration only —
   auto-calibration and the drape mesh pipeline are deleted. A one-shot activation obstruction
   probe (`pickFromRay` on camera activation, clamping the plane short of the first hit) remains;
   ground placement is superseded by the shared-floor v3 behavior below. Calibration persists to
@@ -4442,7 +4444,7 @@ silently demoting every later lookup for the session.
   ground clamp now lifts the cap _center_ only so the wireframe stays a true pyramid welded to
   the plane (was a flattened fan / the ~47.5 m divergence — RESOLVED), re-selecting the active
   camera is a no-op (killed a click-flash), and texture swaps gate on canvas content (killed a
-  periodic white flash). Coverage is now **metro-wide: 250 cameras** (`CCTV_AUSTIN_MAX_SOURCES`
+  periodic white flash). Camera icons are cyan for stills, violet for cameras with a video feed, and amber for the active camera. Coverage is now **metro-wide: 250 cameras** (`CCTV_AUSTIN_MAX_SOURCES`
   default 36 → 250, hard bound 300), filtered to `camera_status === TURNED_ON` (~815 live of
   1,003 rows). City packs (2026-07-04): Caltrans (districts 4/7/11/3 — SF, LA, San Diego,
   Sacramento; cap 300) and TfL London JamCams (cap 250) join Austin (cap 250) as keyless default
@@ -4455,7 +4457,18 @@ silently demoting every later lookup for the session.
   GATHERING stations carry 2,256 in-collection presets, prioritized to 300 against seven anchors
   on the main road spine. `CCTV_MAX_SOURCES` is a 4,000 catalog-wide ceiling shared round-robin
   across packs, so a lower global cap thins every region instead of starving the last pack.
-  ~1,100 cameras total, all RAW PRIOR poses, stills-first.
+  ~1,100 cameras total, all RAW PRIOR poses, stills-first. The catalog lives for 15 minutes; once
+  it expires, `/api/cctv` keeps answering from it while one background refresh reloads every
+  pack, so only the very first load waits for the slowest pack (a live first load took about
+  22 s, and is bounded by Tarktee's 30 s). A background refresh that fails or comes back
+  empty keeps the old catalog and is retried after another 15 minutes, not on the next
+  request. Estonia's Tarktee DATEX feeds, which answer in 2 to 16.6 s, get 30 s
+  (`TARKTEE_FETCH_TIMEOUT_MS`) where other packs get 15 s.
+  Both Tarktee XML catalogs refuse redirects and have a 4 MiB per-document
+  limit, enforced against declared length and streamed bytes. A rejected fetch,
+  unsuccessful response, or body-read failure releases both available response
+  bodies; read failures also cancel an active sibling reader. Cameras sharing
+  an image number are deduplicated before the Tarktee pack cap is applied.
 - **CCTV v3 UX — viewshed + calibration gizmo** (built 2026-07-05 and field
   validated 2026-07-21): the COVERAGE toggle is a
   tri-state cycle `OFF → ON → VIEWSHED`; viewshed mode renders each visible camera's frustum
@@ -4772,7 +4785,7 @@ and unreachable upstream (502/504) separately from road geometry.
   - **Every selectable layer writes that one slot** (2026-08-21). The tracking layers — flights, military, satellites — publish selection on their own awareness lane (`gev:awareness-subject-selected`, consumed by the readout card and the Contacts panel) and for the life of the voice tools never wrote the shared slot, so `scope:'selected'` silently answered `in_view` with a contact plainly selected on screen. They now call `selectTrackedSubjectContext` / `refreshTrackedSubjectContext` / `clearTrackedSubjectContext` (`src/data/contextStore.js`) on select / poll / deselect. **Invariants: (a)** that write path must NOT dispatch `gev:entity-selected` — tracking layers already own an event lane and a second one makes two surfaces fight over one subject; **(b)** exactly one record per tracking layer, because a frozen snapshot of a moved contact must never reach the visible-entity scan; **(c)** precedence is recency, not layer — one slot, last selection wins, so clicking an overlay entity supersedes a tracked plane as the voice subject while the plane stays tracked. Satellites refresh on the 1 s propagation beat, not per frame.
   - **Context-mode vocabulary is symmetric** (2026-08-21). `set_context_mode` accepts `contacts`; the internal id is `flights`. Every model-readable field (`mode`, `entering`, `priorMode`, nested `context`/`contextRollback`, and the transition diagnostic text) is reported in the accepted vocabulary, with the internal id preserved as `<field>Internal`; an absent secondary mode stays `null` rather than claiming to be `off`. Mapping lives in `src/contextModePolicy.js` so UI text and voice payloads cannot drift. Reporting the internal id made the model read `mode:'flights'` as "Contacts is off" and refuse to answer from the `contactsWindow` counts in the same payload.
   - **Analyst → track handoff carries a key, not just a label** (2026-08-21). `analyst_query` items include `icao24`/`mmsi` alongside the display `id`, and contact lookup uses the shared tiered ranking in `src/data/contactMatch.js`: hex exact → callsign exact → registration exact → callsign prefix → registration prefix → callsign substring → registration substring. The tiers keep an exact callsign ahead of a colliding registration regardless of feed order, and registrations compare separator-insensitively (`G-ABCD`/`GABCD`, `05-8152`/`058152`, `N123AB`/`N-123AB`).
-  - **A typed command supersedes the turn it interrupts** (2026-08-21). `sendTextCommand` defers its `response.create` behind an active response instead of colliding with it, marks that response superseded so a late function call from it is refused rather than dispatched, and drops the old turn's queued follow-up. A refused call is still ANSWERED — a terminal `{ok:false, superseded:true}` `function_call_output` — because an unanswered `function_call` strands a pending call and deadlocks the model; the refusal creates no response of its own. A burst of typed commands coalesces into one response while keeping both conversation items.
+  - **A typed command supersedes the turn it interrupts** (2026-08-21; hardened 2026-10-17). `sendTextCommand` defers its `response.create` behind an active response instead of colliding with it, marks that response superseded so a late function call from it is refused rather than dispatched, and drops the old turn's queued follow-up. A refused call is still ANSWERED — a terminal `{ok:false, superseded:true}` `function_call_output` — because an unanswered `function_call` strands a pending call and deadlocks the model; the refusal creates no response of its own. A burst of typed commands coalesces into one response while keeping both conversation items. If the local channel rejects the user item, that command requests no response and adds no deferred answer; earlier accepted deferred input remains queued. Interruption and pointer preparation still happen before the send attempt.
   - **Subject reconciliation across satellite catalog rebuilds** (2026-08-21). A dense↔core toggle or TLE refresh clears and repopulates the catalog, so the published subject is re-resolved against the new satrec; a subject that did NOT survive releases the slot. The per-frame refresh cannot do this itself — `_getTrackedFramePosition` returns early once the satellite has no catalog entry — so the reconcile runs at rebuild completion. An empty catalog is a rebuild in flight, not a disappearance.
   - **One aircraft-proximity snapshot** (2026-08-22; hardened 2026-09-30). `collectAircraftProximityWindow` (`src/data/militaryAwareness.js`) computes the Contacts panel cohort once. The voice analyst reads the exact retained snapshot and evaluation timestamp instead of re-scanning live layers, so the panel readout and spoken count for one centre cannot cross snapshots. **Invariant: do not re-derive a proximity count anywhere else.** Each layer retains 20,000 rows and samples one row beyond the cap; a saturated cohort reports `complete:false`, `truncated:true`, and is spoken as a lower bound. They diverged before because the panel read live billboard positions while the analyst used last-fix coordinates over a 2,000-record slice — 111 on screen, 15 spoken. Explicit regions, arbitrary points, and a radius different from the panel radius deliberately keep the general record engine. Entity-centred results carry `window: {engine:'contacts-window', centeredOn, radiusKm, flights, military, aircraft, complete, evaluatedAt}`.
   - **Centre precedence for nearby asks**: explicit place in the question > Contacts subject (a selected non-contact entity never silently becomes the centre) > an entity the user names > the current view, said aloud. Contacts active with no subject uses the view rather than reading an empty panel.