feat(rail): add a Rail layer with keyless live trains

Live trains from 30 keyless operator feeds: GPS positions from 29 and
Irish Rail's station reports, shown as the operators publish them. 28
work out of the box; Renfe long and medium distance and KTMB appear once
their pack is built, because their vehicles name no route.

Packs are opt-in (scripts/build-rail-paths.mjs): the operator's own
route lines, stop names, next stops and trip-update delays for live
trains. Nothing downloads when the app starts, no train is placed from a
timetable, and no request carries a deployer's key.

SECURITY.md

@@ -74,6 +74,7 @@ The data proxies under `server/providers/` are written so the browser cannot tur
 - **Québec 511 video plays in the browser, opt-in and after Play.** With `CCTV_QUEBEC_VIDEO_ENABLED=1`, `/api/cctv/sources` lists, for each MTQ camera, one URL built server-side from its numeric id on `https://www.quebec511.info/Carte/Fenetres/camera.ashx`; the overlay accepts no other host, path or format. The browser loads that clip only after the viewer presses Play for the selected camera, so Québec 511 sees the viewer's IP address, and the panel says so before Play. Selecting a camera by click, auto-hop, a scene or voice never loads it, and switching cameras stops playback. GEV never proxies, caches or reads the video's pixels. With the flag unset (the default) the browser makes no request to Québec 511.
 - **No verbatim client headers upstream.** The CCTV media route is the one route that relays a request header (`Range`, for video seeking). It is parsed and canonicalized before it is forwarded: one `bytes=` range only, with every accepted form — explicit span, open-ended and suffix — bounded to the same 64 MiB ceiling the relay applies to a declared response body. Multi-range, malformed, inverted and non-`bytes` values are dropped and the request proceeds without a `Range`, as RFC 7233 §3.1 prescribes. Bounding the request bounds what is asked for: a response that declares no length — live streamed media, or a chunked body from an upstream that ignores the `Range` — has no ceiling. An upstream request the browser has stopped waiting for is cancelled rather than left running, whether the viewer leaves before the headers arrive or during the body.
 - **Transit fetches registered feeds only.** `/api/transit/vehicles/<id>` resolves the id against `src/data/transitFeeds.js`; the browser never supplies a URL, and a feed that is registered but disabled does not resolve at all. Redirects are followed manually and each hop is validated against the feed's own https origin before it is requested, so an off-origin or downgraded hop is refused rather than contacted. Bodies are capped at 8 MB, the protobuf is decoded server-side under entity-count and string-length ceilings, a differential feed is refused, and snapshots live 15 s in memory with no disk cache. A per-feed admission limiter and a failure cooldown ladder bound what this process can ask of any operator.
+- **Rail fetches registered feeds only, and none needs a credential.** `/api/rail/feeds` lists the offered feeds; `/api/rail/trains/<id>`, `/api/rail/updates/<id>`, `/api/rail/paths/<id>` and `/api/rail/schedule/<id>` resolve the id against `src/data/railFeeds.js` (`RAIL_FEED_ID_PATTERN`: lowercase letters, digits and hyphens). The browser never supplies a URL, and an id that is not registered, or a trip-updates route for a feed that registers none, does not resolve at all; any method but GET answers 405 (`Allow: GET`, never cached). Every registered feed and every static timetable is open: no rail request carries a credential of the deployer's; one feed's published URL includes a fixed public value its publisher prints for everyone (Bordeaux's `apiKey`), and the only rail setting the server reads is `RAIL_PATHS_DIR`. Upstream requests follow the Transit proxy's policy: https only, each redirect hop validated against the feed's own https origin before it is requested (three hops at most), a request timeout, bodies capped at 8 MiB unless the registry sets a feed's own cap (never above 64 MiB), and a per-feed admission limiter and the failure cooldown ladder bound what the process asks of any operator. Responses are decoded server-side: GTFS-Realtime under entity-count and string-length ceilings, a differential feed is refused, and an empty or header-less GTFS-Realtime reply (an error page served as 200, say) is refused as unreadable, so the last good snapshot is served stale or the request fails with 502. The Irish Rail and SIRI VM XML is read by small scanners, not an XML library: a document that declares a DTD or an entity is refused, and the whole text, the number of trains and every field are capped. Only vehicles the feed's classifier recognises as trains leave the server, and no train position inside Ukraine is ever sent (`src/data/railNoPositions.js`). Snapshots and trip updates live in memory for the feed's own freshness window, with serve-stale after a failure; there is no disk cache and no history. Path and schedule packs are optional build output, read from disk (`.cache/rail-paths/` or `RAIL_PATHS_DIR`) by a reader that re-checks the id against the same pattern, so a name never leaves that directory. A pack over 24 MiB is refused before it is read and answers 503, never a partial file, and so does a pack that cannot be read: only a missing pack is 404; the rest pass through as bytes, through a 64 MiB in-memory read cache keyed by each file's modification time and size, and are served as JSON holding timetables and track shapes only, `no-cache` with an ETag from that time and size (304 when unchanged). A schedule pack of another format version or another feed, or whose window has ended, is not used, and a feed that cannot be read without it answers 503 rather than an empty list. The pack build (`scripts/build-rail-paths.mjs`) runs only when a person runs it, never on app start: it downloads over https only (a redirect to plain http is refused), caps each download in size and time, sends no credential, and refuses an `--out` folder (or a `RAIL_PATHS_DIR` it writes to) inside `src/` or `public/`.
 - **Local receiver feeds are operator-configured and local-only.** `/api/local-receivers/aircraft` reads only the `aircraft.json` URLs in the server's `LOCAL_RECEIVER_FEEDS`; the browser never supplies an address. Each host must pass the shared tap address rule (`src/data/tapAddress.js`: loopback, RFC1918, `localhost`, `*.local`; no link-local, IPv6 or other names), the scheme must be http(s), the path must end in `aircraft.json`, and credentials, queries and fragments are refused. Invalid entries are logged at startup and never fetched. Reads use a 2 s timeout, refuse redirects, cap bodies at 2 MB and share one read per second; responses name each feed by band only and never carry upstream error text.
 - **Voice place search runs only on the operator's Overpass.** `/api/osm/features` answers only when `OVERPASS_UPSTREAMS` is set and never falls back to a public instance. The client sends a curated preset id, a box, a mode and a limit; the server writes and quotes the Overpass query itself, so no client or model text reaches it. Boxes are capped (36 deg² for a count, 4 deg² for a listing), as are the feature limit, the Overpass timeout and memory, and the response bytes read. Each client (IPv6 clients keyed by /64, as on `/api/overpass`) runs one search at a time and two run at once in all, under a per-client rate limit (12/min) and a global one (30/min); a search gets 22 s across every configured endpoint in turn (Overpass itself is told 20 s), which ends before the voice answer stops waiting at 25 s. The route sits behind the cross-site gate below. Answers are cached in memory and on disk (`.gev-cache/osm-features/`, at most 400 entries and 256 MB, kept a day), keyed to the configured instance so a repointed `OVERPASS_UPSTREAMS` never serves the old one's answers; a malformed answer is never cached.
 - **Validated response formats.** Proxied bodies are relayed only in the format the route serves. CCTV frames and the Street View fallback accept JPEG, PNG, GIF, WebP and AVIF only, sent under their canonical type; any other type falls through to the next fallback. The one exception is a source registered as an MJPEG still (the Taiwan Highway Bureau feeds): its `multipart/x-mixed-replace` answer is read only up to its first JPEG part, capped at 2 MB, and sent as `image/jpeg`; a multipart answer from any other camera falls through. A camera whose API needs an operator credential (the City of Ottawa certificate, `OTTAWA_CAMERA_CERTIFICATE`) has a keyless catalog URL; the frame proxy adds the credential only at fetch time, only for a URL inside that camera's pinned path, so `/api/cctv/sources` and `/api/cctv/stream` never carry it. Live CCTV media must match its feed kind (`video/mp4`, `video/webm`, or `multipart/x-mixed-replace` with a boundary) before it is relayed or the camera counts as healthy, and its redirects stay on the registered origin. GBFS bodies must parse as JSON and are re-serialized as `application/json; charset=utf-8`. Camera catalog URLs are pinned by parsed URL — scheme, exact host, no credentials, default port, and a normalized path inside the agency's image directory — and a redirect on a pinned host must stay inside the same directory. Every `/api/cctv` route uses the keyless cross-site checks described below, so cameras stay reachable behind a reverse proxy; a request carrying forwarding headers never reaches the key-bearing Street View fallback and gets the placeholder frame instead, as when no key is configured.