tech-debt: Frankenstein blast radius breadcrumbs (Vsa10k / L3 / 157)

.claude/board/STATUS_BOARD.md

@@ -243,7 +243,7 @@ compile time. Plan: `.claude/plans/callcenter-membrane-v1.md`.
 
 | D-id | Title | Status | PR / Evidence |
 |---|---|---|---|
-| DM-2 | `LanceMembrane: ExternalMembrane` impl with `project()` + compile-time BBB leak test | **Queued** | Resolve UNKNOWN-1 (ShaderSink overlap?) first |
+| DM-2 | `LanceMembrane: ExternalMembrane` impl with `project()` + compile-time BBB leak test | **In progress** | Phase A shipped `9a8d6a0` — `LanceMembrane` struct + `project()` + `ingest()` + `subscribe()` stub. Phase B: full Lance append + version counter pending DM-4. |
 | DM-3 | `CommitFilter` → DataFusion `Expr` translator (`[query]` feature) | **Queued** | — |
 | DM-4 | `LanceVersionWatcher` — tails Lance version counter, emits Phoenix `postgres_changes` (`[realtime]`) | **Queued** | — |
 | DM-5 | `PhoenixServer` — minimal WS server, Phoenix channel subset (`[realtime]`) | **Queued** | Resolve UNKNOWN-2 (which consumers need Phoenix wire?) first |

.claude/board/TECH_DEBT.md

@@ -80,6 +80,114 @@ looks like, any blocking dependencies>
 Cross-ref: <file:line / deliverable D-id / epiphany entry>
 ```
 
+## 2026-04-24 — Frankenstein blast radius on branch `claude/read-claude-md-jh51O` (Vsa10k / L3 / 157 confusion)
+
+**Status:** Open
+**Priority:** P0
+**Scope:** @integration-lead @truth-architect domain:vsa domain:plans
+**Introduced by:** this branch, commits `468357d`, `7a60c42`, `585f8b0`, `489911b` (plus inherited from earlier `dfcf246` / PR #210 / PR #242-#243)
+**Payoff estimate:** plan-doc surgery (1 day) + contract rename already scoped in IDEAS.md (see 2026-04-19 entries)
+
+Session on `claude/read-claude-md-jh51O` ran out of context before cleanup. Next session needs these breadcrumbs.
+
+### Root errors (three clustered, one session)
+
+1. **"L3" misread.** User said "L3" = CPU cache hardware budget (~16 MB). Session interpreted as cognitive-stack Layer 3 and built a fabricated precision-tier architecture on top.
+2. **`Vsa10k BF16` fabrication.** No such type exists. The real types are `CrystalFingerprint::Vsa10kI8` (10 KB) and `Vsa10kF32` (40 KB legacy) — variants of `CrystalFingerprint`, not precision tiers of `Vsa10k`.
+3. **`Vsa10k = [u64; 157]` inherited confusion.** 157-word bit-packed labelled "10K VSA" is the pre-existing error (source: `dfcf246`). Bit-packed never uses 10,000 — uses 16,384 (powers of 2). Only the high-dim numeric 40 KB / 80 KB forms legitimately carry 10K-D framing, and only for grammar ±5 wire bundling. Workspace already filed the rename in `IDEAS.md` 2026-04-19 entries (157 → 256 for binary; Vsa10k* → Vsa16k* for VSA floats); this session ignored the filed correction.
+
+### Blast radius — plans written this session (needs surgery)
+
+| File | Sections | Verdict |
+|---|---|---|
+| `.claude/plans/callcenter-membrane-v1.md` | §§ 14-17 (commit `468357d`) | Mixed. §14 cold-storage + §15 VSA-UDF-dispatch concept good; concrete sizes/type names poisoned. §16 persona-as-function (32 atoms × 16 weights, PersonaSignature 56-bit) good. §17 four-way multiply + ONNX-replaces-Chronos good; tensor-shape annotations poisoned. |
+| `.claude/plans/callcenter-membrane-v1.md` | § 18 (commit `7a60c42`, lines 1139-1224) | **Delete whole section.** Three-tier precision table, father-grandfather compression, CK-safety proof — all built on fabricated `Vsa10k BF16` and L3-as-cognitive-layer. |
+| `.claude/plans/callcenter-membrane-v1.md` | line 822 (edited `489911b`) | Already partially fixed. Still references "Fingerprint<256>" as if distinct from Vsa10k — verify correctness after rename. |
+| `.claude/plans/unified-integration-v1.md` | DU-3 precision-note (lines 196-201, commit `7a60c42`) | **Delete paragraph.** References §18 that must go. DU-3 body (5 UDFs signature) kept but delegates must use canonical `RoleKey::bind/unbind` from PR #243. |
+| `.claude/plans/unified-integration-v1.md` | DU-1 ONNX classifier (`468357d`) | Good core idea. `recent_fingerprints: Tensor[N, 16384]` tensor-shape comment at line 86 needs review against the 16K binary vs 16K-D float substrate split (see IDEAS.md 2026-04-19 CORRECTION). |
+| `.claude/plans/unified-integration-v1.md` | DU-2 Archetype ECS bridge | Good. Mapping table (Entity=PersonaCard / World=Blackboard / Tick=CollapseGate fire) is sound. |
+| `.claude/plans/unified-integration-v1.md` | DU-4 `rationale_phase: bool` (commit `a05979e`) | Shipped correctly to `CognitiveEventRow`. No change needed. |
+
+### Blast radius — source (this session did NOT modify these; flagging for rename sweep)
+
+- `crates/lance-graph-contract/src/grammar/role_keys.rs` — `VSA_WORDS = 157`, `Vsa10k = [u64; 157]`. Originates `dfcf246` (pre-PR #210). Already in rename sweep per IDEAS.md `CORRECTION-OF 2026-04-19 FP_WORDS`.
+- `crates/deepnsm/src/{content_fp,markov_bundle,trajectory,context,encoder}.rs` — all consume `Vsa10k` / `VSA_WORDS = 157`. Added in PR #243 (`c6e69c4`). Follows whichever direction the rename goes.
+- `CLAUDE.md` P-1 section — references `[u64; 157]` and `trajectory: Vsa10k`. Prose; updates after source rename.
+
+### Blast radius — vsa_udfs.rs (commit `585f8b0`)
+
+`crates/lance-graph-callcenter/src/vsa_udfs.rs` (628 lines) has three wrong operations — flagged separately for the canonical-delegation pass:
+
+- `unbind_op`: set-bit-counting on role-indexed slice → returns fraction. WRONG — should delegate to `RoleKey::unbind` + `recovery_margin` (shipped in `79ac8f6` / PR #242).
+- `bundle_op`: implements XOR, named "vsa_bundle". WRONG NAME — XOR is `MergeMode::Xor` (single-writer delta); violates **I-SUBSTRATE-MARKOV** iron rule if used on state-transition paths. Rename to `vsa_xor` (honest) or implement true CK-safe bundle.
+- `braid_at_op`: cyclic word rotation. WRONG — should use `vsa_permute` (PR #209 reference braid).
+
+### Good ideas salvageable (architectural, not in poisoned sections)
+
+Verbatim user framings from this session, recorded so they don't get lost:
+
+1. **Two callcenter modes coexist.** (a) DataFusion polyglot query mode — addresses the 4096 / 0xFFF command space (SPARQL / SQL / Cypher / GQL / NARS); Redis-like UDF access. (b) VSA blackboard orchestration — roles bind work orders into accumulator; CollapseGate XOR flushes into Markov trajectory; new empty ledger starts.
+2. **VSA lazy-buffer pattern.** Bind events → accumulate in VSA → CollapseGate XOR flush → Markov trajectory → scalar `CognitiveEventRow` projection → external consumer. The accumulator is the lazy buffer; CollapseGate is the bell.
+3. **Kitchen analogy.** Work orders bind into VSA (order sheet) → CollapseGate = bell → callcenter/waiter picks up scalar `CognitiveEventRow` → external consumer gets ticket.
+4. **BBB iron rule already in contract source** (`external_membrane.rs:10`): `Self::Commit` MUST NOT contain `Vsa10k`, `RoleKey`, `SemiringChoice`, `NarsTruth`. Any plan proposing these cross the gate is invalid by construction.
+5. **Grammar ±5 wire format is the ONLY legitimate 10,000-D / 10K-D / 16K-D float carrier.** 40 KB (f32) / 80 KB (u8 5-lane) per the IDEAS.md rename plan. Bit-packed fingerprint substrate is separate: `Container<[u64; 256]>` = 16,384 bits = 2 KB, Hamming-only, NOT VSA.
+6. **Chronos → ONNX replacement** (DU-1): full 288-class `(ExternalRole × ThinkingStyle)` product prediction vs Chronos 1D scalar. Grounding sound; tensor-shape details need re-verification post-rename.
+7. **Archetype ECS bridge** (DU-2): `VangelisTech/archetype` sits ABOVE callcenter; Entity = PersonaCard, World = Blackboard, Tick = CollapseGate fire. Thin adapter crate `lance-graph-archetype-bridge`. Mapping table intact.
+
+### Prior art this session ignored
+
+- `.claude/board/IDEAS.md` 2026-04-19 entries: "FP_WORDS = 256 (supersede the 160 plan)" + "CORRECTION-OF 2026-04-19 FP_WORDS = 256" + "REFINEMENT-OF 2026-04-19 CORRECTION-OF FP_WORDS scope". Explicit governance ban: *"There shall be zero occurrences of '10,000-D binary VSA' / '10,000-bit VSA' in any workspace file."* This session wrote multiple such occurrences.
+- `.claude/board/TECH_DEBT.md` 2026-04-19 `FP_WORDS = 157 (not 160)` entry (exists above this row once this appends).
+- `.claude/board/TECH_DEBT.md` 2026-04-19 `VSA substrate renaming: Vsa10k* → Vsa16k* + float framing` entry.
+
+### Archetype / Chronos — breadcrumbs the plan docs do NOT yet carry
+
+Flagging here what's in this session's conversation context but not in `unified-integration-v1.md` or `callcenter-membrane-v1.md`. Budget honesty: recording only what I can attribute to this session's conversation — not reconstructing.
+
+**Archetype — name collision is undocumented.** "Archetype" in DU-2 refers exclusively to the external `VangelisTech/archetype` Rust ECS crate (simulation layer above callcenter). But the user explicitly noted mid-session: *"What is internal is the person archetypes"* — pointing at `crates/thinking-engine/src/persona.rs` (internal VSA-bound cognitive archetypes). These are two distinct objects sharing the word "archetype":
+
+| Sense | Lives in | Role | Covered in plans? |
+|---|---|---|---|
+| External ECS Archetype | `VangelisTech/archetype` crate | Simulation tick layer ABOVE callcenter | Yes — DU-2 |
+| Internal persona-archetype | `thinking-engine/src/persona.rs` | VSA-bound cognitive identity INSIDE substrate | **No** — only contract-side `PersonaCard` (metadata routing) is named, not the internal archetype |
+
+Next session: add an explicit disambiguation paragraph to DU-2 and to `callcenter-membrane-v1.md` § 16 ("Persona as function") noting that `PersonaCard` (contract, metadata) and `thinking-engine::persona::*` (internal, VSA-bound archetype) are different objects; the callcenter sees the former, never the latter. The "32 atoms × 16 weights" formulation in § 16 sits at the metadata/compression layer — whether it corresponds to or is distinct from the internal archetype is an **open question** (I don't have confident attribution for this from this session's conversation).
+
+**Chronos — what's captured vs what's missing.** Replacement rationale ("1D scalar → 288-class `(ExternalRole × ThinkingStyle)` product"; "ONNX infra already justified by Jina v5 ONNX on disk"; "task = classification, not time-series forecasting") is captured in § 17 of callcenter-membrane-v1.md and DU-1 of unified-integration-v1.md. I do NOT hold confident additional Chronos-specific content from this session's brainstorming that could be added without fabrication. Flagging so the next session knows: if the $200 session's memory of the brainstorm contains richer Chronos material, that's new information to add, not something this session held and failed to record.
+
+**Archetype × Chronos interplay — NOT captured anywhere.** The user grouped "archetype and chronos" together as joint brainstorming content. The plans treat them as independent deliverables (DU-1 Chronos-replacement, DU-2 Archetype-bridge). Whether they compose (e.g. Archetype tick driving ONNX-classifier-as-style-oracle at each tick, feeding Blackboard rounds) is **not documented**. Candidate composition: `ArchetypeTick → UnifiedStep → CollapseGate fire → ONNX classifier predicts next (role, style) → next tick's PersonaCard` — but this is my conjecture, not session-attributable, so flagging as an **open design question** for the next session rather than writing it into a plan.
+
+### ONNX > Chronos — what's in plans vs what's missing
+
+§ 17 of `callcenter-membrane-v1.md` (commit `468357d`) holds a 6-criterion "Why ONNX over Chronos" table: **Output** (1D scalar vs 288 logits) · **Task type** (forecasting vs classification) · **Training** (pre-trained vs Lance E-DEPLOY-1 corpus) · **Precision** (style axis vs role × thinking product) · **Infra** (separate model vs `ort` crate justified by Jina v5 ONNX on disk) · **Fit** (partial vs full product).
+
+**Gap:** the table enumerates where Chronos LOSES but not where Chronos WOULD LEGITIMATELY WIN. User flagged "Chronos only useful for XYZ" as a piece of brainstorm content not captured. I do NOT hold session-attributable XYZ content. Reasoning from first principles (NOT session-attributable, flag if reused): Chronos is appropriate when the task is genuinely temporal forecasting — predicting next F-value N cycles ahead, predicting style-drift onset, predicting gate-commit-rate over a rolling window, auxiliary time-series heads on the Lance internal_cold timeline. These would compose with (not replace) the ONNX classifier's instantaneous prediction. The next session should either fill the XYZ from their own brainstorm recall, or explicitly reject Chronos across all use cases.
+
+### Archetype / persona / thinking-style modeling — epiphany candidates not yet in EPIPHANIES.md
+
+§ 16 and § 17 hold modeling insights that sit in plan docs but have NOT been promoted to dated entries in `.claude/board/EPIPHANIES.md`:
+
+- **Four-way multiply = architecture** (§ 17): `(persona 288 × style 36 × stage 2 × learned-dynamics)` ≈ 20 736 × oracle configurations; F-descent IS the automatic architecture search over this space; misaligned configurations are dropped by the CollapseGate predicate. Epiphany framing: *free-energy minimisation over the four-way product = neural-architecture search without an outer optimiser*.
+- **Persona as function** (§ 16): 32 cognitive atoms × 16 weightings = 16^32 addressable persona space, compressed to 56-bit PersonaSignature. YAML runbooks are macro scaffolding for the context loop, not persona identity. Epiphany framing: *persona identity is a coordinate in atom-space, not a YAML artefact; the YAML is a program on the context loop*.
+- **MM-CoT stage split = faculty asymmetry** (§ 17 row): `rationale_phase: bool` maps to `FacultyDescriptor::is_asymmetric()` (inbound_style ≠ outbound_style). Epiphany framing: *MM-CoT rationale-vs-answer split is NOT a new axis — it reuses the existing faculty asymmetry from the contract*.
+
+**Already in EPIPHANIES.md:** `E-DEPLOY-1` (commit `5dbdf25`) captures the Supabase-shape thinking-extension 9-dim joint epiphany. The three candidates above are NOT yet prepended there. Board-hygiene rule from CLAUDE.md requires EPIPHANIES entries in the same commit as the plan additions — this session violated that rule for § 16 / § 17.
+
+Next session action: prepend three dated EPIPHANIES.md entries using the framings above, cross-referenced to § 16 / § 17 / commit `468357d`. Do NOT write additional modeling content I don't hold; if the next session has richer brainstorm recall, let them author from that rather than inheriting my reconstruction.
+
+### Correction plan for next session (P0 order)
+
+1. **Delete** `callcenter-membrane-v1.md` § 18 (lines 1139-1224).
+2. **Delete** `unified-integration-v1.md` DU-3 precision-note (lines 196-201).
+3. **Append** EPIPHANIES.md entry with the three-cluster root-error summary (L3 / Vsa10k-BF16 fabrication / 157 inheritance), referencing this TECH_DEBT row.
+4. **Fix** `vsa_udfs.rs` three operations to delegate to canonical `RoleKey::bind/unbind`, `vsa_permute`, and an honest XOR name.
+5. **Update** `LATEST_STATE.md` § Current Contract Inventory to include `FacultyRole` + `FacultyDescriptor` (added this session in `2a4a245`, never board-logged — same-commit hygiene rule violation).
+6. **Proceed with** the IDEAS.md Vsa10k* → Vsa16k* rename sweep OR scope-limit it per the REFINEMENT entry (grammar / quantum / ladybug allow-list) — that is its own PR, not a cleanup prerequisite.
+
+Cross-ref: EPIPHANIES.md entries needed; IDEAS.md 2026-04-19 rename-sweep entries; PR #242 (`defe928`) The Click categorical-algebraic click; PR #243 (`c6e69c4`) D5 Trajectory; this branch commits `468357d` / `7a60c42` / `585f8b0` / `489911b` / `a05979e` / `2a4a245`.
+
+---
+
 ### Seeded from PRs #204–#211
 
 ## 2026-04-19 — Contract `ContextChain::coherence_at` returns 0 for non-Binary16K variants

.claude/plans/callcenter-membrane-v1.md

@@ -819,7 +819,7 @@ git verb.
 | Dataset class | Content | Crossing BBB? | Queryable? |
 |---|---|---|---|
 | **External / scalar** | Arrow RecordBatch rows from `project()` | Yes (IS the BBB output) | Yes — DataFusion, Supabase FDW, n8n subscribe |
-| **Internal / VSA** | `Fingerprint<256>` = `[u64;256]` cycle fingerprints (L4/L5 speed tier, 2 KB/row); NARS truth vectors, braid offsets. L3 cold tier can promote to Vsa10k BF16 (20 KB, lossless) or RaBitQ-quantized Lance columns (zero-copy ANN). | No — never crosses BBB | Yes — DataFusion + VSA UDFs (see § 15) |
+| **Internal / VSA** | `Fingerprint<256>` = `[u64;256]` cycle fingerprints (2 KB/row); NARS truth vectors; braid offsets. | No — never crosses BBB | Yes — DataFusion + VSA UDFs (see § 15) |
 
 ### Parallels with git cold storage