[joeblau]

Closes #88
Closes #89
Closes #90
Closes #91
Closes #92

Behavior

Validation runs only when data enters the chart: setData, appendData, prependData, constructor config.data, and updateConfig({ data }). The datafeed's normalizeCandles uses the same rules. Nothing in the frame loop changed.

• Invariant: candle fields (time, open, high, low, close) are finite numbers #88, finite OHLC. Non-finite time, open, high, low, or close is dropped. appendData of a bad candle is a no-op: no load event and no redraw. A NaN close can no longer poison every later SMA value.
• Invariant: candle timestamps are unique in the store #89, unique times. Duplicate timestamps keep the last row, including on DataStore.prepend. Rows that are not strictly earlier than the candle already stored are dropped, so the existing series wins ties. A following appendData at that time updates the one bar.
• Invariant: OHLC is consistent (low <= min(open, close), max(open, close) <= high) #90, consistent OHLC. high and low expand to cover open and close. The caller's objects are not mutated. Autoscale then includes the body and the last-price line.
• Invariant: candle volume is non-negative #91, non-negative volume. A negative or non-finite volume becomes 0. A missing volume stays missing on the core path. The datafeed still writes a numeric volume, and now clamps negatives too. MFI stays inside 0–100.
• Invariant: Candle.time is UNIX seconds (not milliseconds) #92, seconds, not milliseconds. A time above 1e11 (year 5138 in seconds, 1973 in milliseconds) logs one console.warn outside NODE_ENV=production. The value is kept. Nothing is divided by 1000.

Tests

1905 tests passed. scripts/check-coverage.mjs reports 100% line, branch, and function coverage. bun run typecheck:tsc and bun run check:invariants passed.

Benchmarks

node bench/suite/run.mjs --samples=7 --sizes=100000, Node 24, median ms/op.

The first cut scanned and copied, so sorted setData went from 0.8122 ms to 0.9959 ms. #108 stores a valid array by reference after one pass:

Case	Before	This PR	#108
100000 setData sorted	1.0430	1.2069	0.6442

Valid setData is faster than before the validation (0.6442 vs 1.0430 in the same back-to-back run; minimums 0.7880 ms and 0.6145 ms). Overhead on valid data is under 5%.