Avoid repeated history materialization in study execution

[mikamikasuki]

Summary

• Poll scheduler readiness using narrow work metadata rather than materializing complete durable payloads and result histories on each scan.
• Read only the pending-work and provider-profile fields needed for eligibility checks; keep original execution payloads untouched until claim.
• Keep the latest full action history in the work checkpoint, while state-transition events store a replace-from-turn delta. Replaying event deltas reconstructs each historical action-history version without copying all prior turns into every new event.

Validation

• Focused runtime, scheduler, context-query, and end-to-end regressions: 118 passed, 1 skipped.
• Added large-history tests that verify readiness and event serialization do not deserialize or repeat cumulative history, while complete history remains recoverable.