Continuous monitoring and early warning for track-record scores

[prasants]

Continuous monitoring and early warning

Stacked on feature/score-firewall (#1); review that first. This adds the
deterioration-detection layer the marketplace-operator workflow needs: a
point-in-time score says how good a record is today, monitoring says whether
it is deteriorating — early enough to act.

Dual-window scoring

• The expanding window is the authoritative published score.
• A trailing window is compared against the preceding equal-length
  window. Equal lengths make the comparison length-fair — a naive
  "recent vs. whole history" gap is contaminated by record length, since
  sufficiency and the Deflated Sharpe reward more data and make every short
  window look like decay.
• Divergence is measured on annualised Sharpe, not the composite: the
  composite saturates at the short-record cap on sub-year windows, which
  would flatten the signal exactly where it must be sharp. TRAILING_DIVERGENCE
  is the pre-emptive alert; it fires before the slow expanding score reacts.

What's in it

• score.monitoring — Cadence (daily/weekly/monthly/yearly), Monitor,
  ScoreSnapshot, ScoreHistory, Alert / AlertKind / Severity.
• Alerts: trailing divergence, new critical flag (manipulation emerging
  over time), grade drop, score drop — each referencing its triggering
  snapshot.
• Auditable history: an append-only, hash-chained ScoreHistory;
  is_chain_valid() proves the record was not reordered or back-edited, and
  each snapshot stays firewall-verifiable.
• Thresholds are operator-tunable parameters; delivery (webhook, dashboard,
  polling) is a hosted adapter over the emitted data structures.

Quality

• 25 new tests; full score package suite (59 tests) passing. ruff lint +
  format clean. Additive: no changes to the scoring engine or firewall.

Docs

• docs/monitoring.md — specification, dual-window rationale, and the
  calibration/delivery boundaries.