v2.1.0

Aggregate state: stateless cohort summary

recon batch --summary emits one aggregate-only cohort summary over a batch:
observability-adjusted prevalence (observed rate, conservative lower bound, and
observability fraction, so missing-not-at-random absence is encoded honestly),
aggregated posterior mass with a separate high-confidence share, and provider and
cloud concentration (entropy and HHI). Add --json for machine output; the
default renders a compact panel. Stateless and compute-and-forget: it ships no
baselines, stores nothing, makes no baseline-relative anomaly score, infers no
unobserved services, and names no domain in its output.

• The summary carries its own record type (cohort_summary, schema_version
  2.1); the locked v2.0 per-domain schema is unchanged.
• Caller-grouped analysis (grouping, distinctive-slug ranking, partial pooling)
  stays in a downstream reducer under validation/aggregate/, which shares the
  per-cohort math with core so the two never drift. See docs/aggregate-state.md
  for the methodology and a fully synthetic worked example.
• --summary pairs with --json or stands alone; it does not combine with
  --md, --csv, or --ndjson.

Gate: full pytest, ruff, pyright (0 errors), validate_fingerprint.