Persist agent memory and isolate prompt evaluation

[MemoryWorld]

Fresh agents skipped their first memory write because empty memory objects were falsey. Persist complete memory snapshots, summary counters and access counts, restore them across runs, and reject stale concurrent writes. Reflexion now restores temporary state on failures and persists its attempt history.

Add separate OPRO tuning tasks, explicit evaluation-protocol metadata, isolated benchmark sessions, and deterministic demo dependencies without global model patches. Include offline regression tests and Windows/Linux CI.

Validation: 53 offline tests passed on Windows/Python 3.12; critical Ruff checks and git diff checks passed; the standalone offline demo completed with provider/network guards covered by tests.

Evaluation limits are documented: the QA scorer remains a substring smoke test, synthetic demo scores do not measure model quality, and model token cost/latency were not benchmarked. Memory and trajectory writes are separate transactions; background job status remains process-local.