Numerically-stable O(1) BufferList reformulations + IsApproximately test helper

[DaveSkender]

Background

PV001 (shipped) found that the naive four-sum identity (Σy² − (Σy)²/n) for Slope.BufferList suffers catastrophic cancellation on stock-price-like inputs and fails the IsExactly bit-equality contract, so Slope.BufferList stays at its current ~3.41x deviation-form ratio for v3.0. These are the follow-up paths.

Sub-items

• Welford/Pébay sliding-window O(1) Slope variant — numerically-stable sliding-window updates via Welford running mean + Pébay M2 deltas avoid the cancellation and may bit-match Series. Prototype against Slope.BufferList test suite; ship only if .IsExactly(series) holds.
• Single-pass Slope with running sumY — less ambitious: eliminate the _buffer.Average() pre-pass while keeping the deviation foreach. Bit-equality not guaranteed; measure drift, then decide whether to accept under a relaxed tolerance or revert.
• IsApproximately<T> test helper + BufferList tolerance audit — cross-cutting decision to allow finance-grade approximate comparisons in BufferList tests where bit-equality is incompatible with O(1) numerical reformulations. tests/Library/TestTools/TestAssert.cs today exposes only IsBetween and IsExactly. Add an IsApproximately(precision) helper, audit all 79 BufferList test files, and decide per-indicator whether IsExactly is load-bearing or IsApproximately(Money6) is the correct contract. Unlocks the O(1) identities above and similar cancellation-bound reformulations elsewhere.

History

Originally tracked as PV-V31-1, PV-V31-2, and PV-V31-3 in docs/plans/streaming-indicators.plan.md (now removed; full write-up in git history).