fix: correct the directional movement seeding off-by-one

[phmatray]

Split out of #110 so a semantic change to shipping indicators gets reviewed on its own, with tests, rather than riding inside a fluent-API PR.

Adx, Dx, PlusDI and MinusDI seeded one term short — the same shape as the EMA seed corrected in #105. The C reference uses while (i-- > 0), which runs its body i times; the C# transcription was while (true) { i--; if (i <= 0) break; }, which runs it i - 1 times.

Two measured symptoms

Both on a 100-bar linear uptrend (high = 100 + 2i, low = 99 + 2i, close = 99.5 + 2i):

1. The output claimed bars that don't exist.

	BegIdx	NBElement	BegIdx+NBElement	bars
Adx	27	75 → 73	102 → 100	100
Dx	14	88 → 86	102 → 100	100
PlusDI	14	88 → 86	102 → 100	100
MinusDI	14	88 → 86	102 → 100	100

Two values claimed for bars 100 and 101. Any consumer mapping an output element onto a bar — which is the documented contract, and what the samples in #104 do — was misaligned by two, silently.

2. The ADX seed averaged 13 values while dividing by 14.

On this series DX is a constant 100, and Wilder's average of a constant is that constant, so ADX must be exactly 100 at every bar. It read 92.857 — which is 100 × 13/14 — then crept toward 100 without arriving:

before:  first=92.8571  last=99.9703
after:   first=100      last=100

Adxr is affected transitively. +DI, -DI and DX values were already correct and are unchanged — only their alignment metadata was wrong.

Tests

The existing AdxTests, DxTests, PlusDITests, MinusDITests assert only RetCode.Success, so none of this was detectable by the suite. The four fixtures here assert numbers, and I verified how each behaves against the old code:

Fixture	vs. old code
OutputNeverClaimsMoreBarsThanTheInputHas	fails (102 ≠ 100)
AdxOnAConstantTrendIsExactlyOneHundredFromItsFirstValue	fails (92.857 ≠ 100)
DirectionalIndicatorsSeparateAPureUptrend	passes — control
DirectionalIndicatorsSeparateAPureDowntrend	passes — control

The two controls are deliberate: they pin +DI = 80, -DI = 0, DX = 100 (all exact and hand-derivable on a linear trend) to prove the fix leaves the already-correct values alone rather than trading one error for another.

without the fix:  Failed: 2, Passed: 2
with the fix:     Failed: 0, Passed: 4

Verification

dotnet build TaLibStandard.sln -c Release  ->  0 errors
dotnet test  TaLibStandard.sln -c Release  ->  1129 passed, 0 failed
  Candles 673 · Functions 236 · Backtesting 220

Release note

This changes the numeric output of Adx, Adxr, Dx, PlusDI and MinusDI in a published package. Callers who compensated for the two-bar misalignment will need to stop. Worth calling out in the v4 notes alongside the ATR/EMA/RSI fixes from #105.

Once this merges, #110 (fluent API) should be rebased on top — it currently carries these same four files, and its strict alignment guard is what surfaced the defect in the first place.

🤖 Generated with Claude Code