fix(dashboard): chart Avg windows flat-line at 0 on wide ranges (#168 downsampling regression)

[VijitSingh97]

The bug

On the 24h / 1w / 1mo ranges, selecting any Avg window other than the default 10 Min — i.e. 1 Min / 1 Hr / 12 Hr / 24 Hr — renders a flat line at zero (with the y-axis then auto-scaling to the tiny Shares-marker range, ~1.4). The 1h range works for all windows.

Reported from gouda:

Range	Avg	Result
1 Hr	1 Min	✅ real data (37–48 kH/s)
24 Hr	1 Min	❌ flat at 0, y-axis maxes at ~1.4

Root cause

build_chart downsamples whenever a range has more points than its canvas tier (_downsample_history, #47). The downsampler bucket-averaged only the legacy v / v_p2pool / v_xvb columns and silently dropped the per-window columns added in #168 (v_p2pool_1m, v_xvb_1m, v_*_1h, v_*_12h, v_*_24h). _split_values then looks up e.g. v_p2pool_1m on the downsampled row, finds nothing → 0, and the v-fallback only rescues the default 10m window. So every non-default Avg window collapses to 0 on any range wide enough to downsample. The 1h range (≤ point tier) keeps native rows, so it's unaffected — which is exactly the observed pattern.

Not missing data — verified on gouda's DB: the last 24h holds 1,896 non-zero v_p2pool_1m points (max 47,681 H/s), yet the chart drew zero.

Fix

_downsample_history now carries every value column through, derived from HASHRATE_WINDOW_COLUMNS so a newly-added averaging window is averaged automatically and can't regress the same way:

_DOWNSAMPLE_VALUE_COLUMNS = tuple(dict.fromkeys(
    ("v",) + tuple(col for cols in HASHRATE_WINDOW_COLUMNS.values() for col in cols)
))
# ...bucket-average each column in _DOWNSAMPLE_VALUE_COLUMNS

Display only — no change to data capture, the algorithm, or the default 10m series.

Tests

• test_downsample_preserves_per_window_columns — asserts the bucket-averaged rows keep every per-window column.
• test_wide_range_keeps_nondefault_avg_nonzero — end-to-end build_chart(24h, avg=1m) over a downsampled series stays non-zero (fails before this fix).
• Full pytest 518 passed; make test-inventory regenerated; drift guard green. The Fake-daemon mini-stack CI job rebuilds + loads the dashboard.

(Answers the "grey out empty selections?" question too: the selections aren't empty — the data was being dropped — so fixing the plot is the right call rather than disabling the buttons. The 12h/24h Avg windows still legitimately need that much uptime to fill, which #168 already signposts.)

🤖 Generated with Claude Code