audit follow-ups: conventions documented, mean unit fixed

[Flotapponnier]

Three fixes from the external sanity audit:

1. perp-open-interest: methodology states the one-side-per-contract convention and names exactly where DefiLlama diverges (Lighter/Ostium doubled long+short per their own adapter code, Paradex includes 1,682 options markets, Hyperliquid includes HIP-3 sub-dexs). Our numbers verified correct against each venue's own API; the footnote pre-empts the "DefiLlama says 2x" objection.
2. pm-data-freshness: anchor rows (Polymarket, Kalshi at ~0.5ms) documented in methodology as the canonical T0 reference, not competing feeds.
3. polymarket-resolution-delay: mean queries multiplied to ms to match the percentile convention (payload mixed ms percentiles with a seconds mean; any API consumer hit the inconsistency).