fix(strategy): make a flat RSI window neutral instead of oversold

[kcrmin]

The divergence

Live and backtest computed RSI with two different implementations that disagreed on a documented edge case.

	flat window (no gains, no losses)
data-pipeline (producer)	50 — "a market that did not move is not a market that only rose"
backtest-engine (consumer)	50 — "It must be reproduced verbatim"
trading-engine (live)	0 ← the outlier

Backtest reads the published RSI_14 series and never recomputes; live computes in-process from closes. So the same strategy was oversold live and neutral in its own backtest.

A flat window has no relative strength to divide, so the value is a convention rather than a result — and both other implementations chose 50 with the reasoning written down. As an indicator bounded at 0 and 100, returning 0 also means maximum oversold, which is the wrong reading of a market that did not move at all.

Why it mattered in the dangerous direction

With an UP crossing at threshold 30:

• old: previous = 0 → sat below the threshold, so a single later rise crossed it → candidate emitted
• new: previous = 50 → never below the threshold → no crossing

Live therefore entered trades the strategy's own backtest never showed. That is precisely the guarantee the official backtest exists to provide, so this was a reproducibility break rather than a rounding nit.

Test

Sixteen closes whose previous window is perfectly flat and whose current window rises once — the minimal shape that separates the two values. It emits a candidate under the old value and none under the pinned one.

Note on scope

Only the flat-window convention is changed. A second, much smaller divergence remains: the pipeline quantizes each running average to eight decimal places at every step, while the live interpreter sums gains and losses and divides once. That can differ in the low digits and only matters for a threshold landing exactly on the boundary. Aligning it means reproducing the pipeline's quantization recursion, which is worth its own change rather than being smuggled in here.

Verification

:modules:strategy-runtime:test --tests "*BasicPlanInterpreterTest*" passes locally (JDK 21).

Context

Found while auditing the strategy domain for business-logic correctness against catalog basic-elements:2026-08-08.

🤖 Generated with Claude Code