Add consolidateFacts() to the MemoryStore tier (TS and Python): distill
selected memories into atomic, deduplicated facts via a caller-provided
extractFacts LLM seam and write each as its own memory, keeping the source
memories so recall is preserved. Facts are reconciled by subject (newest
date wins, tombstones drop a subject) and each fact's asserted date is
preserved in its content.

Off by default; enable via the consolidation constructor flag
(true or { enabled, factSource, factImportance }). Prior fact memories are
excluded from the candidate scan so a re-run never re-distills its output.

reconcile emitted a noop as soon as an incoming fact matched the prior
statement, before comparing dates. A later restatement carrying a newer
date was dropped, so consolidateFacts could persist a stale [date] prefix
despite the documented newest-date-wins rule.

Same-statement restatements now update when the incoming date is strictly
newer (refreshing the stored date) and stay a noop for equal/older/dateless
dates (no needless rewrite). Fixed in TS and Python for parity, with
regression tests in both.

isNewer compared date strings with a `?? ''` fallback, so a dateless
candidate ('') lost to any dated prior. A later dateless assertion with a
different statement was therefore dropped, contradicting the documented
rule that a dateless fact is the latest assertion and should win.

isNewer now returns true when the candidate is dateless (it is the newest
thing we know) and otherwise compares its date against the prior (a
dateless prior counts as the epoch). Fixed in TS and Python for parity,
with regression tests in both.

…ries

consolidateFacts reconciled each batch against an empty existing set, so
re-runs wrote duplicate fact rows and cross-run tombstones/updates never
superseded prior fact memories. Persist the reconcile key (subject) on each
fact memory, load the stored facts for the scope, and reconcile against them:
a re-run over unchanged sources rewrites nothing (idempotent), a newer dated
statement supersedes (deletes) the prior memory, and a tombstone retracts it.

The result now reports created (ids written) and deleted (prior fact memories
superseded or retracted). Writes happen before deletes so a crash between the
two can only leave a duplicate, never lose a fact. TypeScript and Python in
lockstep with mirrored regression tests.

…t dupes

Two correctness edges in write-time fact consolidation (TS + Python mirror):

- [date] statement ambiguity: a dateless fact whose statement starts with
  a bracket (e.g. "[Q3] revenue target is 5M") round-tripped through the
  content regex into a spurious dated fact (date="Q3"), and isNewer's
  string compare then dropped a genuinely newer dated restatement
  ('2' < 'Q'). Persist the fact's date in its own hash field and use it as
  the source of truth for datedness; strip the known "[date] " prefix only
  when a date is present. Content encoding is unchanged, so the diff logic
  is untouched.

- Consolidation race: existingBySubject used map-set, so two concurrent
  consolidateFacts runs that each wrote a fact for the same subject left a
  duplicate that was silently dropped from tracking and orphaned forever.
  Keep one canonical row per subject and retract the extras on the next
  run (self-heal).

Also brace the one-line isNewer guard. Adds regression tests on both sides
(dateless-bracket round-trip, newer-dated-wins, duplicate retraction).

…oncile

factContent/storedFactToFact/buildMemoryRecord already treat date === ""
as dateless, but isNewer and isStaleTombstone only special-cased
undefined. So a fact with date: "" sorted as an empty string (before
every real date): it failed to supersede a dated stored fact, and a
date: "" tombstone looked stale so the subject was not retracted.

Centralize the rule with a factDate helper (empty string -> dateless) and
use it in isNewer, isStaleTombstone, and the same-statement restatement
check, so every comparison agrees with the storage encoding. Mirrored in
the Python package. Adds regression tests on both sides (empty-date
supersede and empty-date tombstone retract).