[BUG] Temporal edge deduplication collapses disjoint occurrences and skips interval finalization

[thantiklermcirony]

Bug Description

An incoming fact with the same text and endpoints as a historical fact can reuse that historical edge even when their supplied validity intervals are disjoint. The new occurrence is lost. The same issue exists in the semantic-duplicate result path and within-batch deduplication.

Environment

Graphiti 0.30.2 at eaa4128681bc53487138a4bbc22d58336ebe70d2, Python 3.12.14, Windows, development checkout. This tests the real resolver with supplied dates and a fixed response at the model seam; it does not evaluate natural-language extraction.

Minimal reproduction

Run with the chosen checkout first on PYTHONPATH and its dependencies installed. No graph service or model account is needed.

import os
os.environ['GRAPHITI_TELEMETRY_ENABLED'] = 'false'

import asyncio
from datetime import datetime, timedelta, timezone
from types import SimpleNamespace
from graphiti_core.edges import EntityEdge
from graphiti_core.nodes import EpisodicNode
from graphiti_core.utils.maintenance.edge_operations import resolve_extracted_edge

t0 = datetime(2026, 1, 1, tzinfo=timezone.utc)
common = dict(
    source_node_uuid='person', target_node_uuid='project', name='ASSIGNED_TO',
    group_id='repro', fact='Alex is assigned to Payments.', created_at=t0,
)
old = EntityEdge(
    uuid='old', **common, valid_at=t0,
    invalid_at=t0 + timedelta(days=10), expired_at=t0 + timedelta(days=11),
)
new = EntityEdge(uuid='new', **common, valid_at=t0 + timedelta(days=20))
episode = EpisodicNode(
    uuid='episode', name='Reassignment', content=new.fact, group_id='repro',
    source='message', source_description='supplied fixture', valid_at=new.valid_at,
)

async def supplied_response(*args, **kwargs):
    assert kwargs['prompt_name'] == 'dedupe_edges.resolve_edge'
    return {'duplicate_facts': [0], 'contradicted_facts': []}

async def main():
    result, _, _ = await resolve_extracted_edge(
        SimpleNamespace(generate_response=supplied_response), new, [old], [], episode
    )
    print('resolved UUID:', result.uuid)
    print('valid_at:', result.valid_at.isoformat())
    print('invalid_at:', result.invalid_at.isoformat() if result.invalid_at else None)
    print('historical end preserved:', old.invalid_at == t0 + timedelta(days=10))

asyncio.run(main())

Actual: UUID old, interval [2026-01-01, 2026-01-11). The January 21 recurrence disappears.

Expected, and observed with the candidate: UUID new, start 2026-01-21, no end; the old January 11 end remains unchanged. A semantic duplicate response must not override known disjoint intervals.

Proposed scope

The correction is limited to temporal bookkeeping in edge_operations.py:

• Guard exact and semantic reuse against separate intervals; retain different supplied intervals in batch deduplication. Preserve the source reference clock while both dates remain unknown.
• Resolve missing dates before historical duplicate eligibility, using the edge's own reference time and at most one timestamp attempt. Do not invent a start from an end or from ingestion time.
• Finalize expired_at for an ended fact even when no candidates exist. Currently an unrelated candidate determines whether that metadata is recorded.
• Clip an older finite interval at a known later contradiction inside it, preserving prior transaction expiration and never extending its end. Currently an already-set expired_at bypasses clipping.

Compatible duplicates and unknown-time controls remain covered. This is not complete interval algebra or a new storage schema. The end-only RELEASED extraction report in #1841 remains unresolved; supplied valid-start fixtures do not establish its end-to-end cause or resolution.

Proposed source · Identity regressions · Resolution regressions

Validation and limits

The two new regression modules contain 33 passing cases on the candidate, covering temporal identity, reactivation, touching/disjoint intervals, batch ordering, expiration, finite clipping and timestamp-reference compatibility. All semantic decisions are supplied fixtures.

Combined-candidate evidence, including the separate Neo4j datetime correction reported in #1864: 430 unit tests passed, 11 skipped; 26/26 declared valid-time query assertions passed across two real query paths, compared with 16/26 on baseline. These cover 13 fixture timepoints on Neo4j Community 5.26.30, not model accuracy or isolated validation of this resolver change. One planned fulltext-dependent database case remained unverified after sandbox index-creation failures; full integration is incomplete.