docs(handovers): BindSpace dissolution architectural delta — from bardioc runtime-session

.claude/handovers/2026-06-05-0445-bardioc-to-lance-graph-bindspace-arch-delta.md

@@ -0,0 +1,153 @@
+# Handover — BindSpace dissolution architectural delta
+
+> **From:** bardioc runtime-session (2026-06-05)
+> **To:** lance-graph session (fresh, when spun up)
+> **Canonical doc:** `bardioc/BINDSPACE_DISSOLUTION_HANDOVER.md` (+ bardioc PR #18) — read that for the full ~440-line delta. This note is the pointer + the headlines per the handovers protocol.
+
+## What I did
+
+- Walked the architectural arc for BindSpace dissolution in a runtime-session
+  conversation (2026-06-05).
+- Captured the durable delta as `BINDSPACE_DISSOLUTION_HANDOVER.md` in
+  bardioc (PR #18) + an append-log entry in bardioc's
+  `CROSS_SESSION_COORDINATION.md`.
+- Pushed this thin handover note here per the `.claude/handovers/` protocol
+  so a fresh lance-graph session lands on it via the standard bootload.
+- Did **NOT** execute the migration plan (this session's domain;
+  out of bardioc's MCP scope + needs lance-graph board-hygiene grounding).
+
+## FINDING
+
+- **Vsa16kF32 has three real loads, two with replacements + one research-only:**
+  - Markov context-chain bundling → **standing wave** (`lance-graph-planner::temporal`, shipped PR #468).
+  - BBB inner/outer ontology helper → **Rubicon** (`bardioc/rubicon/`, shipped bardioc PR #17).
+  - 5^5^5 holograph experiment → `sentence_crystal` + `holograph` crate, **research artifact**, NOT migration scope.
+- **The 16k width is a storage constant**, not a cognition constant
+  (literally one Lance "book" register at 64 KB = 16384 × f32). Cognition
+  has no opinion on the number; storage does. The cycle plane dies (E-BATON-1);
+  the 16k width survives as the book-aligned storage I/O unit.
+- **Post-P64, the columns are already L1-L4 cascade-shaped** (per plan §10.5:
+  64 / 256 / 4096 / 16384 = AVX-512 register / AMX tile row / 4 KiB page /
+  L1d cache). The LE contract isn't *adding* a lossless invariant — it's
+  naming a property the columns already have.
+- **MailboxSoA D-MBX-A1 inherits 4 of 5 BindSpace columns as the same
+  LE-contract types** (`edges` / `qualia` / `meta` / `entity_type` —
+  same `CausalEdge64` / `QualiaI4_16D` / `MetaWord` / `u16`). G5 + G6
+  in the engine_bridge function matrix are **1-line retargets through
+  typed accessors**, not rewrites. **The migration is value-preserving
+  (no conversion), NOT layout-guaranteed byte-identical** —
+  `CausalEdge64` is `pub struct CausalEdge64(pub u64)` *without*
+  `#[repr(transparent)]` in `crates/causal-edge/src/edge.rs:117`, so a
+  slice-reinterpretation pattern (`&[u64] as &[CausalEdge64]` via
+  `transmute`) is **NOT** supported. Use typed accessors throughout
+  (`mb.set_edge(row, e)`, `mb.set_qualia(row, q)`, etc.); the function
+  matrix in the canonical doc §5 already does this. Adding
+  `#[repr(transparent)]` to `CausalEdge64` is a separate concern in
+  `crates/causal-edge/` and would unlock zero-copy slice patterns if
+  ever wanted later.
+- **The 2026-05-27 plan is still authoritative.** Its §3 column map and §6
+  gated steps are correct; this delta narrows scope and adds context.
+
+## CONJECTURE
+
+- Adding the `content/topic/angle` Hamming planes to `MailboxSoA` (still gated
+  on D-MBX-A2) unlocks G1 ingress + G4 cycle-half — that's the gating gap
+  before S1-S4 can ship.
+- **Radix-trie addressing** (PackedDn generalised to variable-length paths)
+  composes with Lance's append-only / standing-wave subtree-scan-at-`v_ref`.
+  OGAR identities (`ogit-erp/sale.order/42`) are already prefix-rich; Rubicon's
+  `LanceMembrane::commit_event` keyed on `inv.object_instance` becomes a trie
+  append, the audit log = `traverse_subtree("ogit-erp/sale.order/42/", *)`.
+- **Qualia trie codebook with Quintenzirkel rotational invariance** is a
+  follow-on D-MBX (compression-only, independent of the singleton
+  dissolution). Frozen-set + circle-of-fifths progression structure →
+  8 B → 1-2 B per row; task-scoped activation via `QualiaScope` analogous
+  to OGAR PR #28's `_effectiveReaders` (tax masks `arc/numbness/*`; invoice
+  masks `arc/fear-of-death/*`; therapy = full codebook).
+- **Layer correction**: `Staunen` and `Wisdom` are PLASTICITY STAGES, not
+  qualia archetypes. They live on `plasticity_counter: [u8; N]` and its
+  derivatives, not in `qualia`. The CLAUDE.md "Staunen × Wisdom qualia"
+  framing was a misnaming (see canonical doc §8). They CORRELATE with
+  qualia-trie-miss on genuinely novel input but neither implies the other.
+- **The trie-miss IS the Staunen signal** — side effect of a lookup you
+  were doing anyway, with structural context (which prefix you missed at)
+  a flat hash collision-set couldn't give. Cheaper than BLAKE3-per-row for
+  novelty detection.
+- **Zero-copy audit** is the residual discipline after S1-S4 ship. Fires
+  at cross-boundary copies, re-encode membranes (canonical bad pattern:
+  `engine_bridge.rs:262-267` f32→i4→column), Arrow translations on hot-path
+  data, DTO-as-payload-when-could-be-view. Does NOT fire at transient
+  bundle compute inside one think-arc (E-BATON-1 carve-out).
+- **64 MB working set fits one TiKV region** (default ~96 MB, splits at
+  ~144 MB). The deferred cross-server HLC tick gets its natural home —
+  Raft log = tick sequence, region leadership = single-writer authority
+  for tick advancement, region splits = the only event needing
+  reconciliation across leaders. No custom protocol.
+
+## Blockers
+
+- **D-MBX-A2** (add Hamming planes + ratify temporal/expert decisions on
+  `MailboxSoA`) is the gating gap. S1-S4 cannot land until A2 ships.
+- **No active lance-graph session today** — this handover restores state
+  when one starts.
+
+## Open questions
+
+- **Bucket 2** (`lance-graph-cognitive::storage::bind_space::BindSpace`,
+  ~20 sites across `cypher_bridge.rs` / `spo/sentence_crystal.rs` /
+  `fabric/{executor,zero_copy}.rs` + `learning/{scm,feedback}.rs`) is a
+  SEPARATE migration arc per `docs/BINDSPACE_MIGRATION_GAP.md` (the
+  ladybug-style Container/CogRecord/PackedDn port). Confirm before
+  treating it as singleton scope.
+- **Bucket 3** (`sentence_crystal` + `holograph` crate) is the 5^5^5
+  research artifact (cube-with-Schnittpunkten geometry mentally /
+  sandwich layers in IT reality). Stays as research; no action needed
+  unless someone tries to make it a carrier (then I-VSA-IDENTITIES fires).
+- **Qualia trie codebook timing** — independent of the singleton
+  dissolution; can ship before or after. The 8 B → 1-2 B compression rung
+  earns its keep when the codec-stack rotation makes room.
+- **StreamDto / ResonanceDto / BusDto DTO collapse** (plan §2.6) is
+  mostly mechanical post-S2; worth a doc pass when the time comes.
+- **D-MBX-7** (lance-graph container layout = MailboxSoA layout =
+  `ndarray::simd_soa.rs` aligned) per plan §11.1 stays as the
+  prerequisite for the §2.7 D-MBX-6 (zero-copy SurrealDB view).
+
+## Sources
+
+- **Canonical** (read first): `bardioc/BINDSPACE_DISSOLUTION_HANDOVER.md`
+  (bardioc PR #18).
+- **The plan** (still authoritative): `.claude/plans/bindspace-singleton-to-mailbox-soa-v1.md`.
+- **Cross-session ledger**: `bardioc/CROSS_SESSION_COORDINATION.md` —
+  2026-06-05 append-log entry mirrors this handover.
+- **Prior art**: `docs/BINDSPACE_MIGRATION_GAP.md` (bucket-2 ladybug arc),
+  `CLAUDE.md` "The Click" + 2026-05-26 Baton scoping note (Vsa16kF32
+  framing, deprecated as a carrier).
+- **Rubicon binding** (the BBB replacement): `bardioc/rubicon/` (Phases
+  1-5 shipped, bardioc PR #17 merged); `bardioc/MIGRATION_SPINE.md` for
+  the spine view; `bardioc/STANDING_WAVE_ARCHITECTURE.md` for the
+  deinterlace engine (= the standing-wave replacement for Vsa16kF32's
+  Markov-bundling load).
+- **OGAR docs** affected: none — the contract surface does not change.
+  `QualiaScope` (if/when it lands) is a class-level annotation analogous
+  to `ContextDepends`; no breaking change.
+
+## Punch list for execution (if you pick this up)
+
+1. Ratify **D-MBX-A2** — add `content/topic/angle: [[u64; 256]; N]` Hamming
+   planes to `MailboxSoA<N>`; decide on `temporal` (fold into
+   `CausalEdge64.temporal` or keep `[u64; N]`) and `expert` (subsume into
+   `mailbox_id` or keep optional column).
+2. Ship **S1** — G5 + G6 retargets in `engine_bridge.rs`
+   (`write/read_qualia_*` + `persist_cycle` minus cycle plane). 1-line
+   edits behind a feature gate; both `&BindSpace` and `&MailboxSoA`
+   overloads alive in parallel.
+3. Ship **S2** — G4 dissolve. Replace `dispatch_busdto`/`unbind_busdto`/
+   `busdto_to_binary16k` with a `BusDto::view_over(&MailboxSoA, row)` +
+   inverse. `engine_bridge.rs` shrinks ~600 → ~150 LOC.
+4. Ship **S3** — G1 reshape. `ingest_codebook_indices` becomes a thin
+   sensor adapter per plan §2.6 StreamDto fate.
+5. Ship **S4** — delete `Arc<BindSpace>` from `ShaderDriver`; drop the two
+   `BindSpace::zeros(4096)` allocations in `bin/serve.rs:29` +
+   `bin/grpc.rs:29`. Singleton dissolved; plan §5 reached.
+6. Then: zero-copy audit pass per the canonical doc §9; D-MBX-A6 (planner
+   DTO overhaul) per plan §11.6.4; optional qualia trie codebook follow-on.