…E, supervision-leg over-claim caught

Retroactive epiphany-brainstorm-council gate (the gate skipped when the three
identity epiphanies were hand-prepended this session). 5-savant panel, every
claim re-verified against source. Verdict: 3x REVISE, no REJECT.

- E-ANCESTRY-TRINITY-1: the 'supervision-edge = is_ancestor_of, same relation'
  leg is NOT wired -- is_ancestor_of has zero call-sites in lance-graph-supervisor
  and orchestration.rs (grep-verified). Demoted identity -> CONJECTURE so a future
  session does not delete supervisor routing believing the bit-shift subsumes it.
  Legs 1-3 (one-shift subClassOf, template specialization) stay code-true.
- E-MINT-TRACE-1: present-tense bug -- 'registry NOT deduped' was true at
  write-time but D-IDENTITY-2 (5aaa54c) shipped the dedup the same session; the
  :476 mint cite went stale (now 565-579). Two-mints-coexist caveat added
  (canonical registry mint vs legacy positional contract/ontology.rs:85, gated
  by deferred D-IDENTITY-3).
- E-OGAR-NORTHSTAR-1: DECISION stands; provenance pointer added.

Applied append-only: prepended E-COUNCIL-IDENTITY-1 + updated the 3 Status lines
(past entry bodies untouched). AGENT_LOG entry for the council run.

https://claude.ai/code/session_014A4JuRCqKP2yNENrQ9Ha7H

…iberate append-only exception)

User asked to drop the two embarrassing identity epiphanies; chose the full
sweep, so this also removes the council-verdict entry that narrated them.

Removed from EPIPHANIES.md: E-MINT-TRACE-1 (tense bug — its 'not deduped'
claim was inverted by D-IDENTITY-2 the same session), E-ANCESTRY-TRINITY-1
(unwired supervision-leg over-claim), E-COUNCIL-IDENTITY-1 (the retroactive
gate verdict on the two). E-OGAR-NORTHSTAR-1 (a ratified DECISION) and
E-IDENTITY-WHITEBOX-1 are kept; E-OGAR's back-references to the removed
entries were stripped so it stands alone.

De-dangled all cross-references: STATUS_BOARD epiphany list, two AGENT_LOG
work-record entries (kept the substance, dropped the removed-ID pointers),
and removed the council-run AGENT_LOG entry. Zero residual references remain.

This is a deliberate one-off exception to the append-only board rule, made
on the repo owner's explicit instruction; the audit trail is this commit.
The underlying code (D-IDENTITY-2: dedup + bijection, 5aaa54c) is unchanged
and remains correct — only the ledger prose was removed.

https://claude.ai/code/session_014A4JuRCqKP2yNENrQ9Ha7H

…taFusion UDF + Cypher over one HHTL address space

Research deliverable: mailboxes parse as a normal cold path because (a) the
fork's KV keys are storekey-encoded order-preserving and record-id ranges
lower to native byte-range scans, so an HHTL subtree under the left-aligned
addr64 codec is ONE contiguous range (is_ancestor_of = range containment);
(b) the mailbox needs only the 5-method Transactable read subset over a
phase-pinned SoaEnvelope snapshot to be indistinguishable from Lance cold.
The ontology registry (entity_type<->NiblePath bijection, D-IDENTITY-2) is
the resident Christmas-tree skeleton; rows materialize at their address from
whichever tier owns them; DEFINE TABLE/FIELD round-trips through the C16b
DDL builders as the catalog exchange format.

Grounded by two parallel repo sweeps + main-thread spot-verification; one
agent claim corrected (SoaEnvelope has zero real impls -- identity N3 stands,
now D-PG-2). kv-lance found FULLY implemented in-tree (surreal_container's
BLOCKED C/D stale). LanceDB-leading ruling respected throughout: SurrealQL
is a dialect/view, never the store.

Board: INTEGRATION_PLANS prepend, STATUS_BOARD D-PG-1..6 (Queued), AGENT_LOG.

https://claude.ai/code/session_014A4JuRCqKP2yNENrQ9Ha7H

…c foveated tree-builder

User direction confirmed against identity.rs octet layout: the GUID's left
half (classid + HHTL address) is order-preserving plain bytes, so Cypher
label/subtree patterns compile to byte-prefix range predicates servable by
Lance zone-maps/scalar indexes; quantized similarity (RaBitQ-style/CAM-PQ/
Binary16K) rides the same row. Two caveats: namespace sorts first (cross-
namespace template scans = per-ns ranges or registry hop); GUID carries only
4 path nibbles (deeper scans via registry/addr64).

New M6 + D-PG-7: NiblePath assignments computable by a deterministic
hierarchical partitioner (CLAM pole-split preferred over order-dependent
classic Louvain), 16-way capacity-bounded so foveation falls out (depth grows
where dense), and APPEND-STABLE as an iron requirement -- bootstrap once,
greedy insertion after, minted paths never move (tree = address space;
re-clustering would re-address every GUID). Determinism + append-stability
property tests mandatory.

https://claude.ai/code/session_014A4JuRCqKP2yNENrQ9Ha7H