OHLC candles for one coin

HL `candleSnapshot` window. `interval` ∈ { 1m, 5m, 15m, 1h, 4h, 1d }. `limit` clamped to [1, 5000]; default 100. Cached 30s (the most recent candle is still forming — clients refreshing more often than that get the same in-flight value). Historical reach: bars newer than the HL node's earliest indexed fill are live-aggregated; older bars come from a write-through cache. The top 14 HL crypto majors (BTC, ETH, SOL, NEAR, TON, WLD, XRP, SUI, ZEC, TAO, DOGE, ADA, AVAX, BNB) have a 2-year Binance Futures backfill at 1m/5m/15m/1h, so windows past HL's own 5000-bar reach (5m ≈ 17d, 1m ≈ 3.5d) still return data for those names. HL-native names (HYPE, xyz:* HIP-3 perps, @N spot deploys) have no Binance counterpart and remain capped at HL's 5000-bar window. Each returned candle carries a `source` tag (`hl_node` | `binance` | `official_api`) so backtesters can audit provenance per-bar — `null` only on responses from older adapter builds. Three windowing strategies — pick whichever fits the caller: 1. Neither `start_time` nor `end_time` → most recent `limit` candles (default). 2. Both `start_time` + `end_time` → all buckets in `[start_time, end_time)`, capped at `limit` newest. 3. Only `start_time` → from that timestamp forward, capped at `limit`. 4. Only `end_time` → ending at that timestamp, the most recent `limit` candles. `start_time` and `end_time` accept either Unix-ms (number) or an ISO-8601 string (e.g. `2026-05-19T00:00:00Z`).