feat(integrity): per-bench expected_n + sample-health badge

[Flotapponnier]

Bench pages currently rank providers based on sample counts that
vary from 3 to 1500+ per provider depending on the bench cadence,
without surfacing the confidence level. The methodology page
advertises n >= 1000 which is misleading on low-cadence benches
like oracle-deviation (3 to 4 samples per provider) and perp-funding
(3 samples per venue by design).

This change introduces a per-bench expected_n declared in the YAML
spec. The page computes health = actual / expected_n per provider
and badges rows below 50 percent of expected as "low sample" or
hides rankings below 10 percent as "insufficient". The bench-wide
aggregate (median of per-provider healths) also propagates to
/api/citable, /api/stat and the headline sentence so an "insufficient"
field stops asserting a winner in machine-readable feeds.

YAML field is optional. Benches without expected_n declared behave
as before (no badge, no hide).

YAMLs touched with expected_n:

• oracle-deviation: 4 (source count per pair, sample_size query
  counts distinct sources not data points)
• perp-funding: 3 (BTC + ETH + SOL asset series per venue; the
  sample_size query counts distinct asset series)
• aggregator-head-lag: 50000 (~5760 per chain-region cell, 9 cells
  per provider summed by the sample_size query)
• bridge-quote-latency: 3000 (4 routes x 3 amounts x 12 sweeps/hour
  x 24h is ~3456, conservative)
• bridge-fee: 1000 (4 routes filtered to amount_usd=300 x 12/hour
  x 24h is ~1152, conservative)

YAMLs left WITHOUT expected_n (cadence is heterogeneous or the
sample_size semantics differ per provider, no honest single value):

• validator-yield (Solana reports ~200, Hyperliquid ~30)
• l1-finality (cadence varies per chain: WS continuous vs 10s
  poll, blocks per chain differ by 1-2 orders of magnitude)
• l2-block-time (block cadence ranges from Arbitrum 250ms to zk
  rollup bursts of 6-30s)
• network-fees (no sample_size query declared)
• gas-estimation (cadence per oracle differs)
• stablecoin-peg (CEX 5s tickers vs Curve 12s on-chain mixed)
• buyback-audit (7d window, sample count not meaningful per scrape)
• network-coverage (12h freshness, dimension count not data points)
• metadata-coverage / nft-collection-metadata / wallet-labels-coverage
  (coverage benches, sample_size is dimension count)
• rpc-capabilities (capability flags, no time-series sample count)
• hyperliquid-frontends + hip3-deployers (per-builder gauges,
  builder count not data point count)
• pm-* benches (cadence varies per provider)
• token-deployment-cost (per-chain attempt)
• polymarket-resolution-delay (event-driven, not periodic)
• perp-fees (per-asset per-venue, count varies by support)
• stablecoin-peg-usdt-anchored (mirrors stablecoin-peg)

Thresholds chosen: 0.5 (healthy), 0.1 (insufficient). See
src/lib/sample-health.ts.

Test plan:

• pnpm validate (31 specs valid)
• pnpm typecheck
• pnpm lint
• bun test src/lib (105/105 pass, includes the new
  sample-health.test.ts)
• pnpm build
• Visit /benchmarks/oracle-deviation on the staging Preview URL:
  expect XRP/ADA/DOGE to read as healthy (3/4 = 75 percent) but the
  badge surfaces once sources drop to 1 of 4.
• Visit /benchmarks/aggregator-head-lag: expect no badge
  (healthy across all 3 providers).
• Hit /api/citable for oracle-deviation, perp-funding: expect
  dataConfidence + expectedN fields in the JSON.