feat(dashboard): price earnings history from the platform's own figure

[flyworker]

Earnings over time was computed locally — this node's token counts times published per-million rates. Those rates are not the whole of how a provider's earnings are determined, so the chart was an arithmetic exercise denominated in dollars, sitting next to the platform's actual total and inviting the two to be compared.

What the platform actually offers

I probed it before designing anything. There is no earnings time series:

• /provider/me/earnings, /earnings/history, /stats/history, /stats/daily, /history, /payouts, /settlements, /rewards → all 404
• /provider/me/stats returns lifetime totals and ignores ?period=, ?from=/to=, ?granularity=, ?days= — every variant returns the identical total

The approach

That lifetime total is cumulative and monotonic. Recording it with each metrics snapshot and differencing consecutive samples gives what the platform says was earned in between — no local rate arithmetic at all. It also sidesteps the restart problem the local counters have, because the platform's figure never resets.

Costs one upstream call every couple of minutes at most; the stats client already caches.

Honest about what is still an estimate

• Intervals with no recorded figure — stored before this, or sampled while the API was unreachable — keep the local estimate. Each point carries authoritative, the series carries authoritative_points, and the footer states the provenance instead of presenting both as the same number.
• The split by model stays local. The platform reports no per-model breakdown, so the share is rescaled onto the authoritative total to fill the bar. The bar's height is the platform's number; how it is divided is this node's estimate, and the UI says exactly that.
• A decrease contributes zero — a lifetime total going down is a correction or a payout on the platform side, not negative earnings.

Verified

9 tests, including: the ledger delta is used even when a deliberately absurd local rate would give a wildly different answer; a bucket with no platform figure does not claim to be authoritative; the rescaled model segments sum to the bar and preserve the local 3:1 ratio; a decreasing total yields zero.

8/8 packages, go vet and gofmt clean, tsc clean, rebuilt bundle, config documented.

Worth flagging

If the platform updates that figure on a settlement cycle rather than continuously, differencing will show flat stretches then jumps. That is truthful, but it changes the chart's shape — and a real time-series endpoint on the platform side would be the better long-term answer than sampling a running total.