The published candleset was a filtered copy of Binance, not a reproduction, and three
independent defects did it silently.
1. parser.py deleted every candle Binance itself stamped off its timeframe grid. Binance
re-phased its kline clock during a few maintenance windows (BTCUSDT 1m steps 21 s
forward on 2017-12-04 and runs at that phase for two weeks; BTCUSDT 1h opens at :28:15
for a few hours on 2018-02-11), so 65,716 published rows never reached the dataset --
20,403 of them from one BTCUSDT-1m-2017-12 archive alone. BinanceVisionRequest gains
timestamp_grid_policy: 'keep' retains those rows as bounded WARNING evidence, the
default 'strict' keeps the old deletion for an unverified feed.
2. _format_number rendered every field through '.15g', which is not round-trip safe. Binance
quote volumes carry 16-17 significant digits (1198158182.23741980) and were flattened to
1198158182.23742, so every published row disagreed with the venue on quote_volume and
taker…
Replace the entire commit history with a single root commit so that a full clone plus \git lfs fetch --all\ costs 42.9 GB instead of 104 GB.
The tree is byte-identical to the previous tip 4d4ee6d (aa207e35), so no
LFS object is re-uploaded.
Branches removed from origin in the same cleanup:
2026less f506091 (37.4 GB of exclusive LFS history, 18.8 GB tip)
2026only 64ebea1 (3.1 GB of exclusive LFS history, 3.1 GB tip)
