reconcile compared each incoming fact only to the initial existing snapshot,
so same-batch duplicate subjects, supersession chains, or a tombstone after
an add produced wrong ops. Fold each op into the working map as the batch is
walked so later facts see earlier add/update/delete results.

A fact restated identically across sessions kept only the first session's
session_id (reconcile NOOP), so its fact chunk could carry a session_id absent
from answer_session_ids and miss recall. Track every session asserting a fact's
current statement and emit one fact chunk per source session.

Wrap the JSON.parse in parseFacts so a malformed LLM reply (or prose with
stray brackets) degrades to no facts for that session instead of throwing and
aborting the whole eval run. Export parseFacts and cover it with unit tests.

- reconcile ignores a stale tombstone dated before the curated fact (mirrors
  the supersede staleness check); a dateless tombstone still deletes
- consolidateRecordFacts tags each per-session fact chunk with that session's
  own date instead of the first assertion's, so restated chunks order correctly

isNewer treated a missing candidate date as the empty string, so an
undated later update always lost to any dated prior and the stale fact
won. Sessions are reconciled chronologically, so a dateless candidate is
the later assertion — give it the same winning default a dateless
tombstone already has.