fix(runtime): publish position metrics under the precision rules

[kcrmin]

Paired with backtest-engine's counterpart. Together they make the two runtimes publish the same number for the same position.

position.returnPercent and the peak and drawdown beside it were rounded HALF_UP at scale 8. precision:1.0.0 is HALF_EVEN at scale 8, and the official backtest quantizes these HALF_EVEN accordingly — so an exact tie at the ninth decimal published two different numbers for one position.

These are the values a POSITION_RETURN, PEAK_RETURN or DRAWDOWN_FROM_PEAK step compares against its threshold, and their rendered form reaches the step trace. So the disagreement was both in what the strategy decided at the boundary and in what the two runtimes recorded having decided.

Testability

PositionTracker becomes package-private, as ExecutionGate beside it already is, so the metrics can be tested directly rather than only through the whole evaluation loop.

PositionTrackerTest pins:

• the rounding — fails against develop with expected: <0.00000002> but was: <0.00000003>
• the peak carried across bars, so a fall from it is a drawdown rather than a loss
• the entry bar counting as the first held bar
• trading days being elapsed rather than inclusive, which is why the catalog can offer 당일 장 마감 as a distinct option from a one-trading-day hold

Verification

./gradlew build — BUILD SUCCESSFUL, whole repo.

🤖 Generated with Claude Code