feat: add peer strategy evidence exchange

[douglas-grishen]

What

• add a public @agentplat/collective-runtime/strategy-evidence-exchange API for signed, content-free peer outcome attestations
• enforce exact strategy, implementation, feedback, cohort, membership epoch, and configuration bindings
• add bounded causal state, duplicate/replay handling, out-of-order retention, atomic pending-chain draining, certificates, and handoff
• quarantine equivocation and protect per-peer heads against instance/stream churn
• add conservative robust aggregation across independent peers and source groups
• integrate collective evidence as an optional, confidence-scaled prior while preserving local baselines and vetoes
• provide CollectiveSync and sparse-overlay gossip adapters without introducing a runtime dependency cycle
• document the architecture, security model, public API, and implementation status

Why

Peers need to learn from independently observed outcomes without exchanging sensitive task content, accepting remote control, or depending on a global coordinator. This capability keeps collective information advisory and leaves every adaptation decision under local policy authority.

Safety properties

• Ed25519 signatures and canonical digests
• current membership and local trust eligibility
• fail-closed schema and state validation
• bounded storage and bounded influence
• durable causal heads and authenticated epoch rotation
• equivocation quarantine with contribution removal
• local baseline floor and veto preservation

Validation

• full workspace build passed
• full workspace type-check passed
• public audit: 2,497 files, 0 secret/restricted matches
• unit suite: 864 tests, 857 passed, 0 failed, 1 skipped, 6 todo
• packed-consumer verification: 48 tarballs and 95 public API surfaces

Review note

This PR is intentionally stacked on the bounded local strategy adaptation branch so the evidence prior integration can be reviewed incrementally.