A failure now ends when its city leaves range or the camera rises above the
activation gate, so an empty map no longer reads LOAD FAILED. Late activation
and status rejections are ignored like late successes. Partial failure reads
DEGRADED, every in-range city failing over cached points reads STALE, and
freshness advances only on a refresh that succeeded. Behavioural tests drive
the real proximity, update and enable paths. Docs say no licence is published
for Toronto and list Social Bicycles and PBSC. The merge also restored the #61
GBFS v3 changelog entry that an earlier merge on main dropped.
The layer recorded failure in a single `_error` string. Every city feed is
fetched independently, so one retired operator set that flag for the whole
layer — and the only code that cleared it ran on a successful *activation*,
which `runProximityCheck` skips for a city that already has rendered points.
A dead feed that re-entered camera range therefore set the flag on every
camera move and nothing ever cleared it: a permanent LOAD FAILED over a map
drawing every other city's stations.
Failure is now tracked per city in `_failedCityIds`, and `getStats()` reports
it only when the failures left nothing to draw. A city that recovers removes
its own entry rather than waiting for a neighbour to succeed.
This was reachable only because Buffalo sits 95 km from Toronto, inside the
100 km activation radius — the one dead city in the registry with a healthy
neighbour in range. The other four failed alone, where LOAD FAILED with an
empty map reads as correct, which is why the latch went unnoticed.
Probing all 33 registry systems found five dead. Three had moved and are
repointed: Reddy Bikeshare (Buffalo) to Social Bicycles, Tugo (Tucson) and
Trinity Metro (Fort Worth) to PBSC, all verified serving stations. Two are
dropped because the systems wound down rather than moved — Broward B-cycle
404s at every documented endpoint, and CoGo's replacement Lyft feed publishes
zero stations with status last reported in March 2025. `gbfs.cogobikeshare.com`
no longer resolves, so its allowlist entry goes too.
All 31 remaining systems were verified live, and every registry URL was
checked against the proxy host and path allowlists.
Toronto runs the same PBSC platform as the Austin, Honolulu and
Chattanooga systems already in the registry, and publishes the identical
GBFS v2 endpoint pair under `toronto.publicbikesystem.net`. The proxy
allowlist already admits the host through its `*.publicbikesystem.net`
wildcard and the paths end in the expected `station_information.json` /
`station_status.json`, so the system needs no new plumbing — only the
registry entry that names it.
Verified against the live feeds: 1,069 stations in both documents, all
1,069 joining on `station_id`. The layer reads `num_bikes_available`,
which Toronto populates normally, so availability renders without a
vendor-specific mapping. It does not populate the standard
`num_bikes_available_types`, reporting per-type counts under PBSC ids
(ICONIC/BOOST/EFIT/ASTRO) instead; nothing reads that field today, and
this entry deliberately does not translate it rather than adding a
mapping with no consumer.
The center is downtown at 43.6532, -79.3832 with the group's 90 km load
radius. The stations span roughly 37 km east-west, from Etobicoke to
Scarborough, so a tighter radius would drop the outer ends on a downtown
pan.
The new registry test is the reason this was cheap to be sure about: it
holds every city's feed URLs against the same host and path predicates
the proxy enforces, so a city the proxy would silently refuse now fails
the suite instead of rendering an empty layer.
Credited to Bike Share Toronto (Toronto Parking Authority). The City's
open-data package declares no licence and the feed publishes no
`license_url`, so it is treated as attribution-only, like the other GBFS
operators, rather than claiming a licence that was never granted.
Assert through computeInRangeCities that a camera over each system's
outermost live-feed station activates it, including both Kona and Hilo
for HIBike, and that HIBike stays off over Honolulu. Drop Pennsylvania
from the changelog: Philadelphia's Indego is already registered.
Register four live US systems on Lyft's PBSC platform: MoGo (Detroit, 81
stations), POGOH (Pittsburgh, 60), WE-cycle (Aspen, 101) and HIBike (Kona
and Hilo, 23). The proxy already allows *.publicbikesystem.net, so no
server change is needed. HIBike's centre sits between Kailua-Kona and
Hilo so a camera over either town loads the whole system.
The test checks each system is registered, passes the proxy's host and
endpoint checks, and has an activation radius covering its farthest
station as measured against the live feed.
Cover every Norwegian station-based city-bike system with an open GBFS
feed: Kolumbus Bysykkel (Stavanger) and Farte Bysykkel (Skien/Porsgrunn)
through Entur Mobility, and Bysykkel Lillestrøm through JCDecaux
Cyclocity.
api.entur.io and api.cyclocity.fr join the GBFS allowlist, each kept to
its GBFS path prefix since both hosts front other APIs. Entur serves its
feeds without a .json suffix, so the endpoint checks accept either form,
and requests to Entur carry the ET-Client-Name header it requires.
Attributed under NLOD (Entur) and Licence Ouverte (JCDecaux).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Add Norway's three Urban Sharing city-bike systems to the Bike Share
layer and get_bike_share. The feeds are keyless GBFS on
gbfs.urbansharing.com, which joins the proxy allowlist. The proxy now
sends the Client-Identifier header the operators ask clients to send.
Attributed under NLOD 2.0 in DATA_SOURCES.md and the data credits.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keeps the voice-core prompt rules and adds main's world-query and
vessels-vs-visuals rules; the tool catalog wraps the deixis-aware runner.
Also fixes two things the branch had before the merge:
- Live voice could not start: OpenAI rejects track_entity's top-level
anyOf. Model-facing tools now omit top-level combinators; the runner
and benchmark still require a query or a referent.
- The Contacts-centred analyst answer dropped the viewport coverage note.
The bike-share layer parsed station documents inside its model factory.
Move the parsers unchanged to src/sources/gbfsStations.js and have the
model import them, so code outside the layer can read GBFS stations
without constructing the layer.
- 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.
Supersedes #218 by @spcpza, ported onto the current component layout with fixes from review.
Verified before merge: format, package boundaries, unit suite, build, and the browser tracking suite; a browser harness for the change where one applies; adversarial review applied.
Contributor: @spcpza. His original commit is the base of this branch, and his own re-port onto main (2026-09-15, `5b69011`) was consulted while finishing the height work. That port reads every vehicle's surface from the app's shared ground-floor system (`cachedGroundFloor` / `warmGroundFloor` / `GROUND_FLOOR_LIFT_M`) rather than sampling the scene per layer, keeps a vehicle hidden until its cell answers, warms a bounded batch per poll on a timer instead of per frame, and renders at the ellipsoid above the altitude gate. That is the sounder primitive — one floor system for the app, with the rooftop-versus-street question already settled against a datum-correct DEM prior — and this branch's height work was rebuilt on it, replacing a bespoke cell cache, ground prior and rooftop rule written here. Credit for that direction is his.
Future work (not in this PR), both written up in docs/CURRENT-STATE.md:
- Readable route names. The realtime feeds carry only `route_id`; `route_short_name` / `route_long_name` live in each operator's static GTFS `routes.txt`, and headsigns in `trips.txt`. Doing it properly wants a server-side metadata loader beside the transit proxy with fixed registered static sources, bounded indexes built once and persisted, a slow conditional refresh, and a versioned route dictionary the browser caches. Shipped here: ids are never rewritten, and a cut one carries an ellipsis instead of reading as a typo.
- Below-ground rendering. Subway vehicles are drawn on the surface directly above themselves, because the feeds carry no depth and a train nobody can see is not worth drawing; the card says so plainly. Real treatment wants OSM tunnel geometry per city and a ghosted sprite tethered to the surface. The owner recorded it and left it.
Round four and after (2026-09-15). The owner tested round three under FLIR and NVG, and then asked for the design to be taken further: elegant in performance, beautiful on camera, trails with a cache for public transport, Cesium-native, and a core other moving layers can adopt. Codex Astra reviewed the plan (it confirmed the speed fault: a glide started from the drawn position at receipt time but spanned the gap between report times, so with MBTA reporting every ~10 s against a 15 s poll every second glide ran at 1.5x; it rejected a queued-segment patch), then wrote the design, then implemented it in commit-sized passes with adversarial review between them and real-browser harness runs by the integrator.
What this PR now contains beyond the contributor's original port:
- Delayed playback in a contact-agnostic core (`src/data/contactPlayback.js`): each vehicle keeps a typed, bounded fix history and a display clock that runs at one second per second through it, a feed-derived lag behind real time, holding at its newest fix rather than extrapolating and never sprinting to catch up. Marker, course, motion word, card and trail all read the same displayed sample. The clock has an explicit time-offset and seek API so a rewind is a later feature, not a rewrite; flights and vessels are untouched.
- A selected-vehicle trail (`src/data/contactTrailRenderer.js`, `src/layers/transit/trails.js`): the retained history drawn with batched Cesium primitives, clipped at display time so the head is exactly the sprite, faded by age, in the mode colour (white under sensor presets), on the same surface path as the marker.
- Bounded proxy-side history (`server/providers/transitHistory.js`): fifteen minutes and at most 128 fixes per vehicle with hard byte ceilings, MBTA only for now (opt-in per feed), served by a small endpoint that never fetches upstream, so a selected trail survives a browser reload while the server keeps running.
- Scene work bounded to what is visible and moving, an occluder, cached DETECT candidates, holds on the render loop only while something visible moves.
- Sensor presets: white-hot haloed sprites under NVG/FLIR/noir (black-hot inverts), saturated mode colours under CRT, per-mode bracket tiers in every theme, the selected card keeping its mode colour.
- A discriminating browser harness (`scripts/qa-transit.mjs`) that samples postRender frames, reads the real card anchor, and measures fleet budgets at 800 and 3,000 vehicles.
Not built, recorded as follow-ups: curated route display names (lowest priority per the owner), fleet-wide tails, route snapping, a rewind UI, adoption of the playback core by flights and vessels.
Verified at merge (2026-09-15 morning): unit suite 3,896 passing, build clean, format and boundary gates clean, tracking regression 109/109, browser harness keyless 200 pass / 0 fail (13 sensor checks unexercised because the live view offered fewer than six isolated sprites; every judged sprite passed), keyed on Google 3D tiles 198 pass with the per-frame allocation gate as the only failure. Follow-ups recorded: the keyed build's playback sampler still allocates per frame (CPU and frame time stay within budget; the 3,000-vehicle metro view sits at a 27.6 ms frame interval against a 25 ms budget), the shared detection overlay's per-paint allocation, friendlier route names, fleet-wide tails, a rewind UI, and adopting the playback core for vessels and flights.
Co-authored-by: spcpza <[email protected]>
Supersedes #441 by @MiguelGFerreira, ported onto the current component layout with fixes from review.
Verified before merge: format, package boundaries, unit suite, build, and the browser tracking suite; a browser harness for the change where one applies; adversarial review applied.
Co-authored-by: MiguelGFerreira <[email protected]>
Supersedes #235 by @cora-fresh-labs, ported onto the current component layout with fixes from review.
Verified before merge: format, package boundaries, unit suite, build, and the browser tracking suite; a browser harness for the change where one applies; adversarial review applied.
Future work (not in this PR): hand-drawn marks do not persist across a reload today, and neither do spoken annotations; there is no storage layer for marks. The natural home is Scenes: a drawing that belongs to a shot and persists per shot, so a scene animation can carry its own marks. Needs an owner.
Co-authored-by: Clint Borle <[email protected]>