cognitive-shader-driver: emit CE64 from the SPOFC record (land #1294 on main)

.claude/board/INTEGRATION_PLANS.md

@@ -1,3 +1,14 @@
+## 2026-09-25 (1) — ce64-spofc-learning-v1 — CE64 as the mantissa that carries SPOFC evidence between cycles → `.claude/plans/ce64-spofc-learning-v1.md`
+
+**Status:** PROPOSAL; leg A shipped (draft PR stacked on #1293). The driver now
+packs each emitted CE64 from its candidate's SPOFC record: palette-space S/O
+instead of a row id, SPOFC truth, the Pearl mask as the inverse of
+`edge_to_layer_mask` (the raw `predicates & 7` turned SUPPORTS into
+CONTRADICTS), and topology bits 59–60 written for the first time in the live
+path. Legs B–E (evidence identity, owner write-back, cross-cycle revision,
+completions) are blocked on named dependencies — the W4a writer, a per-edge
+evidence-identity carrier, a predicate palette — and are not substituted.
+
 ## 2026-09-19 (1) — waben-fold-execution-loop-v1 — from the merged folds to one addressed execution loop → `.claude/plans/waben-fold-execution-loop-v1.md`
 
 **Status:** PROPOSAL. No code authorized; this is the grounded implementation

.claude/board/STATUS_BOARD.md

@@ -1,3 +1,17 @@
+## D-CSL — CE64 as the mantissa that carries SPOFC evidence between cycles (2026-09-25)
+
+Plan: `plans/ce64-spofc-learning-v1.md`. Predecessor: #1293 (SPOFC candidate aggregation).
+
+| D-id | scope | status | gate / falsifier |
+|---|---|---|---|
+| **D-CSL-A** | emit CE64 from the SPOFC record: palette-space S/O, SPOFC truth, Pearl mask as the inverse of `edge_to_layer_mask`, topology bits 59–60 by relation kind | In PR (#1294) | `driver::tests::{emitted_edge_carries_the_p64_target_and_its_evidence, content_partner_edges_do_not_claim_a_direct_topology, causal_mask_round_trips_through_p64_layers}` |
+| **D-CSL-A2** | characterise `learn`: repeat observation recounted; `ALL_FROZEN` emission plasticity | In PR (#1294) | `tests/ce64_recount_probe.rs` — flips deliberately with D-CSL-D |
+| **D-CSL-B** | per-edge evidence identity (stamp) so a repeat or shared root is not fresh support | Queued — blocked (no carrier; W-slot is a corpus root, CE64 bits closed to new semantics) | a revision with the same observation must not raise confidence |
+| **D-CSL-C** | write emitted edges back through the owner path | Queued — blocked on W4a | `persist_cycle` BLOCKED→W4a (`COMPONENT-MAP.md` §6) |
+| **D-CSL-D** | cross-cycle revision of the stored edge with the cycle's SPOFC observation | Queued — after B, C | `ce64_recount_probe` flips |
+| **D-CSL-E** | completions ("Sudoku autocomplete"): inferred edges marked `Indirect*`, one-way probe only | Queued | per `mul-ewa-trust-propagation-v1.md` §0b: trust must not inflate from thinking longer |
+| **D-CSL-F** | Shannon entropy over a candidate's support distribution | Queued | new code; none exists |
+
 ## D-RPT — ReportPlan / zero-copy pivot / report-as-OGAR-projection (2026-09-23)
 
 Entry: `entries/2026-09-23-report-plan-zero-copy-pivot-docir-convergence.md`.

.claude/board/SUPERSESSION-INDEX.md

@@ -14,7 +14,7 @@
 ## What this table says
 
 **`BindSpace` is the shape of the problem.** Marked RETIRE, and simultaneously the
-most-referenced symbol here: **69 crate files, 48 plans, 42 of them blind.**
+most-referenced symbol here: **69 crate files, 49 plans, 43 of them blind.**
 That is a programme, not a cleanup.
 
 **`GateState` is the sharpest case: 1 plan names it and *every one* is blind.**
@@ -52,19 +52,19 @@ coverage instead.
 | `A2AMessage` | BLOCKED | — | 2 | 1 | 0 |
 | `StepMask` | BLOCKED | — | 3 | 10 | 4 |
 | `commit_to_l4` | BLOCKED | — | 2 | 2 | 0 |
-| `dispatch_busdto` | BLOCKED | — | 3 | 8 | 5 |
-| `persist_cycle` | BLOCKED | — | 11 | 8 | 5 |
+| `dispatch_busdto` | BLOCKED | — | 3 | 9 | 6 |
+| `persist_cycle` | BLOCKED | — | 11 | 9 | 6 |
 | `CognitiveMarkers` | REPURPOSE | `Commit` | 1 | 0 | 0 |
 | `DominoCascade` | REPURPOSE | `Commit` | 7 | 1 | 0 |
 | `GateDecision` | REPURPOSE | — | 25 | 28 | 25 |
 | `GateState` | REPURPOSE | — | 14 | 1 | 1 |
 | `MergeMode` | REPURPOSE | — | 8 | 14 | 13 |
 | `ResonanceDto` | REPURPOSE | `PerturbationDto` | 2 | 12 | 7 |
-| `BindSpace` | RETIRE | — | 69 | 48 | 42 |
+| `BindSpace` | RETIRE | — | 69 | 49 | 43 |
 | `CollapseGateEmission` | RETIRE | — | 5 | 14 | 12 |
 | `ThinkingStyle` | RETIRE-toward-contract | — | 52 | 28 | 24 |
 
-## Table 2 — plans naming a ruled symbol without citing the ruling (76)
+## Table 2 — plans naming a ruled symbol without citing the ruling (77)
 
 Route is **mechanical triage, not a verdict**: `ARCHIVE?` = the plan's own status says
 it shipped; `RESCOPE` = it targets a symbol marked RETIRE; `READ` = neither signal fires
@@ -112,6 +112,7 @@ a licence to act on it.
 | **RESCOPE** | `alpha-reason-witness-shader-field-archaeology-pass-1` | `BindSpace`, `MergeMode`, `ResonanceDto` | SOURCE AUDIT / PLAN ONLY. No production wiri | 0/1 |
 | **RESCOPE** | `bindspace-mailbox-soa-dependency-map-v1` | `BindSpace`, `dispatch_busdto`, `persist_cycle` | MAP / preflight. No source wired yet. Read-b | 0/2 |
 | **RESCOPE** | `bindspace-mailbox-soa-w3-w4a-impl-v1` | `BindSpace`, `dispatch_busdto`, `persist_cycle` | v2 — 5-consolidation + 3-brutal-critic pass  | 0/1 |
+| **RESCOPE** | `ce64-spofc-learning-v1` | `BindSpace`, `dispatch_busdto`, `persist_cycle` | PROPOSAL + first leg shipped (draft PR, stac | 0/7 |
 | **RESCOPE** | `cognitive-substrate-convergence-v3` | `BindSpace`, `CollapseGateEmission`, `MergeMode` | ACTIVE — sprint-12 Wave F + Wave G complete  | 4/16 |
 | **RESCOPE** | `mailbox-cycle-aware-write-contract-v1` | `BindSpace`, `dispatch_busdto`, `persist_cycle` | CONJECTURE / design. 5+3-gated before code. | 0/1 |
 | **RESCOPE** | `unified-integration-v1` | `BindSpace`, `MergeMode`, `ThinkingStyle` | Active — brainstorm phase complete; delivera | 0/0 |
@@ -157,6 +158,6 @@ a licence to act on it.
 | **RESCOPE** | `unified-soa-rubikon-integration-v1` | `BindSpace` | — | 0/8 |
 
 - **ARCHIVE?**: 0
-- **RESCOPE**: 56
+- **RESCOPE**: 57
 - **READ**: 20
 - ruled symbols tracked: 14

.claude/plans/ce64-spofc-learning-v1.md

@@ -0,0 +1,125 @@
+# ce64-spofc-learning-v1 — CE64 as the mantissa that carries SPOFC evidence between cycles
+
+**Status:** PROPOSAL + first leg shipped (draft PR, stacked on #1293). Legs B–E are
+blocked on the dependencies named in § Missing dependencies; none of them is
+substituted here.
+
+## Overview
+
+CE64 was meant to keep what a cycle learned for the next cycle. In the live
+dispatch path that role is dormant. #1293 made stage [3] collect each row's
+supporting relationships into one SPOFC record (support count, best partner,
+predicate union, `TruthU8`). This plan carries that record the rest of the way:
+into the emitted CE64, back into the store, and into the next cycle's revision —
+without recounting evidence the store already holds.
+
+## Checklist
+
+- [x] **D-CSL-A. Emit CE64 from SPOFC** (this PR): palette-space S/O, SPOFC truth,
+      Pearl mask as the inverse of `edge_to_layer_mask`, topology by how the
+      relation was found. Disable-verified tests.
+- [x] **D-CSL-A2. Characterise `learn`** (this PR, `tests/ce64_recount_probe.rs`): it
+      pools a repeated observation as fresh evidence; the driver's emission
+      plasticity (`ALL_FROZEN`) stops it moving S/P/O.
+- [ ] **D-CSL-B. Evidence identity.** Decide what tells "same evidence again" from "a
+      new witness" when a stored edge is revised. Blocked: see § Missing, item 2.
+- [ ] **D-CSL-C. Write-back through the owner path.** Blocked on W4a (item 1).
+- [ ] **D-CSL-D. Cross-cycle revision.** After B and C: revise the stored edge with the
+      cycle's SPOFC observation; the probe in A' flips deliberately.
+- [ ] **D-CSL-E. Completions ("Sudoku autocomplete").** One-way probe only, per
+      `mul-ewa-trust-propagation-v1.md` §0b; inferred edges marked by topology.
+- [ ] **D-CSL-F. Support entropy.** Shannon entropy over a candidate's support
+      distribution, as a spread signal beside `m`. New code; nothing like it exists.
+
+## Verified facts this rests on
+
+CE64 layout (default `causal-edge-v2-layout`, `causal-edge/src/layout.rs`, const-asserted
+to cover all 64 bits once):
+
+| bits | field |
+|---|---|
+| 0–23 | S, P, O — three palette256 indices |
+| 24–31 / 32–39 | NARS frequency / confidence, u8 |
+| 40–42 | Pearl causal mask |
+| 46–49 | inference mantissa (signed i4) |
+| 50–52 | plasticity (`0` = `ALL_FROZEN`, `plasticity.rs:17`) |
+| 53–58 | W-slot: witness **corpus** root handle, 0..=63 |
+| 59–60 | `CausalTopology`: Direct 0, IndirectKnownIntermediates 1, IndirectUnknownIntermediates 2, Unknown 3 |
+| 61–63 | `ReasoningBand` |
+
+- **`pack` defaults bits 59–60 to 0 = `Direct`.** Every edge emitted before this
+  PR claimed a direct relation, including content-similarity partners. The two
+  bits share their ordinals with `TrustTexture`; the last writer wins.
+- **The 24-bit S/P/O can carry the target.** The emission simply never wrote it:
+  it packed a row id into palette space (`s = row%256`, `o = (row/4)%256`).
+- **`predicates & 0b111` was not the inverse of `edge_to_layer_mask`**
+  (`p64-bridge/src/lib.rs:61`): plane 2 is SUPPORTS, but Pearl bit 2 reads back as
+  CONTRADICTS. A supporting relation round-tripped as a contradiction.
+- **`learn` (`edge.rs:731`) pools unconditionally**: `c = ws/(ws+1)` with
+  `ws = w_self + w_obs`, no source identity. Revising twice with one observation
+  counts it twice (pinned in A').
+- **Nothing in `run` writes an edge back.** `persist_cycle`
+  (`engine_bridge.rs:784`) writes only `emitted_edges[0]`, has test-only callers,
+  and `ShaderDriver.bindspace` is `Arc<BindSpace>`. The only stored-edge read that
+  crosses cycles is the cascade query `backing.edge(row).s_idx()`.
+- **`tables.revise(...)` is computed and discarded** in stage [3]
+  (`_revised_truth`).
+- **Confidence convention matches arm-discovery; frequency does not.** SPOFC
+  `c = m·255/(m+k)` is `evidence_confidence_u8`, shared with `arm_to_truth_u8`.
+  SPOFC `f` is the best resonance; arm-discovery's `f` is a conditional ratio
+  `cooccur/antecedent`. There is no antecedent count in a dispatch, so the ratio
+  is not available; the difference is recorded, not papered over.
+
+## The "Sudoku" framing, stated against the code
+
+- **8× support = 8× evidence.** In SPOFC terms a row supported eight ways has
+  `m = 8`, `c = 226/255`. That is the evidence a completion is constrained by.
+  It is only honest if the eight are independent; the pre-pass credits each
+  content pair to both rows, and cascade hits through one shared palette target
+  are counted per row. Independence is item B.
+- **Bits 59–60 as the grid folding itself.** Topology is the one place a CE64
+  says how it was obtained: `Direct` (observed plane edge), `IndirectKnown`
+  (a completion whose intermediates are named — its derivation), `IndirectUnknown`
+  (a fill with the path unknown), `Unknown` (similarity, no causal path). An
+  inferred completion must land as `Indirect*`, never as `Direct`, so a later
+  cycle cannot read it back as an observation. This PR is the first writer of
+  bits 59–60 in the live path (Direct / Unknown only; nothing is inferred yet).
+- **Oberflächenspannung / EWA.** `jc::ewa_sandwich` (`crates/jc/src/ewa_sandwich.rs`,
+  contract twin `sigma_propagation::ewa_sandwich`) is SPD covariance push-forward,
+  `Σ' = M·Σ·Mᵀ`. It propagates uncertainty *shape* and counts no evidence. Its
+  only epistemic use is the PROPOSED `mul-ewa-trust-propagation-v1.md`, which
+  names the circularity: Σ → trust → gate → Σ. A fill by a minimisation principle
+  is "plausible everywhere, grounded nowhere", so a completion from it must never
+  raise confidence as if observed. The learned prior it could supply is the
+  surface; the evidence stays the SPOFC count. The driver's "Kerbl EWA"
+  (`alpha_front_to_back_composite`) is scalar alpha compositing and is unrelated.
+- **Shared upstream.** Two supports that trace to one root are one witness
+  (`lance-graph-contract/src/fusion.rs` anti-alchemy law; `shared_roots`,
+  `inherited_roots`). `nars::belief::revise_at` pools only on disjoint, non-empty
+  stamps and otherwise falls back to CHOICE; `admit_derived` admits derivations
+  with an empty stamp so they never pool. That is the discipline item B needs;
+  where its carrier lives in the dispatch path is the open question.
+
+## Missing dependencies (named, not substituted)
+
+1. **The W4a owner write path.** `COMPONENT-MAP.md` §6: `persist_cycle` /
+   `dispatch_busdto` are BLOCKED→W4a, the batch writer pairing
+   `cast(on_behalf = mailbox_owner())`. A write-back that bypasses it violates the
+   V3 ownership rule. Item C waits for it.
+2. **An evidence-identity carrier per stored edge.** `learn` has none. The W-slot
+   (bits 53–58) is a witness *corpus root*, not a per-observation source set, and
+   `le-contract.md` forbids new awareness semantics in CE64 bits. `nars::Stamp`
+   exists in the planner, not in the driver's store. Where the stamp lives (a SoA
+   lane beside the edge, per V3) is a design decision this plan does not make.
+3. **A predicate palette.** P stays 0 because nothing assigns a palette index to a
+   predicate. Using the p64 plane index would be inventing one.
+4. **A production feed into the planes.** `update_planes` has no production caller;
+   `edges_to_layered_rows` has none outside p64-bridge, and `convergence.rs`
+   addresses `%64` where p64-bridge uses `/4`. Closing the topology loop waits on
+   that being reconciled.
+
+## Effects of leg A
+
+Emitted edges change for every dispatch (S/O address space, confidence, Pearl
+bits, topology). `cycle_fp` and top-k are untouched: the braid and ranking read
+the candidates, not the edges.