G3 — Retrieval v2, invalidation triggers, checkpoint parity
===========================================================
tests/memory/test_retrieval.py: 3 passed
tests/memory/test_invalidation.py: 6 passed
tests/runtime/test_checkpoint_parity.py: 2 passed
tests/agentic_state/test_memory_blackboard.py: 5 passed (unchanged)
tests/contracts/test_agentic_memory.py: 3 passed (unchanged)

Proven: deterministic per-signal ranking orders query_memory identically;
semantic bonus is additive and optional. Advisory plans flag stale
environments, touched provenance/contract dependencies and explicit outcomes
without mutating the store. The parity test resumes from a freshly loaded
checkpoint only — decisions, facts, unresolved items, tool path and result
match the continuous run exactly.
