feat(graph): add bi-temporal fact search

[DivyamTalwar]

Why

JITMIND already writes t_observed, t_valid, and t_invalid onto extracted graph relations, but no public graph API could apply those timestamps during retrieval. Worse, a repeated (head, relation, tail) could overwrite the prior relation observation, erasing the history needed for time-travel answers.

What changed

• stores each extracted relation as an immutable, idempotent observation keyed by source memory + triplet
• adds typed GraphFact evidence with source memory/page and full temporal interval
• adds GraphMemoryStore.query_facts(...) with entity, relation-type, namespace, valid-time, and observation-time filters
• adds GraphRetriever.search_temporal(...) so temporal facts flow through the existing Hit contract
• normalizes timestamps before executing parameterized Cypher and bounds result limits
• documents valid-time versus observation-time semantics

Competitive rationale / provenance

Graphiti publicly models dynamic memory as episodes and facts whose valid_at / invalid_at fields can be filtered during fact search: https://github.com/getzep/graphiti and https://github.com/getzep/graphiti/blob/main/mcp_server/src/graphiti_mcp_server.py

This PR implements that missing capability using JITMIND's own Memory, Entity, and RELATION schema and adds an explicit observation-time gate. No Graphiti code was copied.

Evidence

• python -m pytest -q → 42 passed
• python -m compileall -q jitmind → pass
• new tests verify normalized bi-temporal query parameters, namespace and invalidation predicates, stable-but-versioned relation IDs, invalid-input fail-fast behavior, and retrieval provenance

Evidence gap

The Cypher contract is tested through a deterministic driver boundary. A live Neo4j latency benchmark is intentionally not claimed because this runner has no authoritative production graph dataset or Neo4j deployment.

Review notes

This is an independent proposal targeting main; it does not depend on PR #27. Please keep it open for review and comparison.

[DivyamTalwar]

Verification checkpoint

• Exact head: c15ea9bc592136c3b9f8cfdbc3fae4031357ae23
• GitHub Actions CI run: 32663099122 — success
• Matrix: Python 3.10 (97252122494), 3.11 (97252122517), and 3.12 (97252122401) all passed install, pytest, and compileall
• Local suite: 42 passed

Semantics reviewed: relation observations are idempotent per source-memory triplet, later observations preserve history, valid intervals are start-inclusive/end-exclusive, and observation time is independently gated. Live Neo4j latency remains an explicit evidence gap; no performance claim is being made. This PR remains intentionally open and unmerged.