[Epic] Graph model v2 — bi-temporal relational facts

[StigNorland]

Tracking issue for the graph rework designed in paper/graph-model-v2.md (committed 86347bc).

Core: a single bi-temporal Fact edge primitive — (subject)—[predicate]→(object) carrying event-time validity (t_valid/t_invalid) + ingestion time (t_created/t_expired) + provenance. Adopted from Zep/Engram rather than reinvented.

Settled design decisions (doc §10):

• Variable-precision TimeValue (year→subsecond) with recurrence + anchor.
• Open predicate vocabulary; cardinality inferred from data, not a hardcoded table.
• Only concepts merge; entities and facts are singular. Coreference = link, not fusion.
• TIME_POINT kept for date-as-content / co-reference (distinct from date-as-validity).
• polarity (asserted/negated) on facts; opinions = facts whose subject is the holder.
• One unified "referring-expression → node" resolver; local coref resolves structurally, only genuinely-underspecified refs defer to LLM.
• Retrieval gains an as-of point-in-time filter.

Sub-workstreams (to be split into issues when scheduled):

• TimeValue type + variable precision + recurrence
• bi-temporal Fact edge on graph.py Edge
• supersession + inferred per-predicate cardinality (§7)
• entity↔entity (subject,predicate,object) extraction
• polarity + opinion (holder-as-subject) + INSTEAD_OF event relation
• unified referring-expression resolver (entity/time/pronoun coref, LLM-deferred for underspecified)
• as-of retrieval filter
• bottom-up concept/community formation (doc §6 — depends on the above)

Absorbs/supersedes: #15 (signature identity — reframed: concepts merge, entities match), #17 (L2 META gets a job via bottom-up concepts). Blocks meaningful re-eval (#14) once reconnect chores land.

[StigNorland]

Baseline validates the v2 premise. Smoke run shows the current concept layer collapses to 3 hub nodes via over-aggressive signature-merge (details in #16). Confirms v2's stance: only concepts merge, but the merge criterion must be far more conservative and semantically grounded (not Jaccard over shared entity-neighbours). Also confirms temporal QA = 0.0 → bi-temporal + as-of retrieval is the highest-value sub-workstream.

[StigNorland]

v2 progress — foundation + first two behavior-changing steps landed and measured on LoCoMo sample 0 (199 QA, graph mode):

	overall	cat2 temporal	cat3 inference	cat5 adversarial
pre-fix (collapsed graph)	0.293	0.041	0.365	0.830
+ merge fix #16	0.331	0.054	0.538	0.968
+ step4 event-time/as-of	0.336	0.068	0.558	0.968
(baseline mode)	0.286	0.000	0.404	0.968

Commits: eee55eb TimeValue (§2.1) · b36963a bi-temporal Edge (§1,§5–8) · 190099e merge collapse fix (#16) · b350c7b event-time + as-of (§8).

Graph mode now beats baseline on 4/5 categories. Remaining v2 work, in leverage order:

• cat2 temporal (still ~floor) — two levers: (1) surface per-edge t_valid in traversal instead of the concept's creation-turn date (data already stamped); (2) query-time as-of extraction + relative-date arithmetic to actually drive the §8 filter (mechanism landed).
• Supersession (§7) — conflict detection on (subject, predicate) w/ inferred cardinality; untouched.
• Multi-hop regression Multi-hop (cat4) regressed 0.132→0.104 after merge fix #19.

Done so far is committed + tested (tests/test_temporal, test_edge_bitemporal, test_merge_conservatism, test_as_of_retrieval all green).