fix(marginfi-client-v2): cap PythLegacy confidence interval at MAX_CONFIDENCE_INTERVAL_RATIO

[latent-9]

What

For OracleSetup.PythLegacy banks the price confidence interval is effectively not capped. parseOraclePriceData passes PYTH_PRICE_CONF_INTERVALS (2.12) as the cap ratio to capConfidenceInterval, instead of the MAX_CONFIDENCE_INTERVAL_RATIO (0.05) that every other oracle path uses.

Why

capConfidenceInterval(price, confidence, maxConfidence) returns min(confidence, price * maxConfidence), so the third argument is the fraction of price the confidence is capped at. The PythLegacy branch passes PYTH_PRICE_CONF_INTERVALS:

const pythConfidenceRealtimeCapped = capConfidenceInterval(
  pythPriceRealtime,
  pythConfidenceRealtime,
  PYTH_PRICE_CONF_INTERVALS // 2.12
);

PYTH_PRICE_CONF_INTERVALS is the standard deviation multiplier already applied to the raw confidence a few lines above; it is not a cap ratio. Using it as the cap makes the ceiling price * 2.12 (212 percent of price), which the confidence never reaches, so the interval is uncapped for legacy Pyth banks. Every other branch in this function (PythPushOracle, SwitchboardV2, SwitchboardPull) passes MAX_CONFIDENCE_INTERVAL_RATIO (0.05).

The effect is a wider low and high price band than intended. Example at price 100 with a published confidence of 8, so the scaled confidence is 8 * 2.12 = 16.96:

• current: min(16.96, 100 * 2.12) = 16.96, lowest price 83.04
• fixed: min(16.96, 100 * 0.05) = 5, lowest price 95.00

So the SDK values PythLegacy collateral about 12 percent below what the 5 percent cap gives, and over weights liabilities symmetrically.

Change

Pass MAX_CONFIDENCE_INTERVAL_RATIO as the cap in both PythLegacy capConfidenceInterval calls, matching the other oracle paths. The PYTH_PRICE_CONF_INTERVALS multiplier on the raw confidence is unchanged.