docs: cognitive-fabric census 2026-08-30 — SOURCE FACT rows for #1078 W0 (BeliefArena dedup payable; alpha/atom/mask membrane located)

.claude/board/EPIPHANIES.md

@@ -1,3 +1,77 @@
+## 2026-08-30 — E-A-FLOOD-THROTTLE-IS-NOT-A-DISCRIMINATOR-1 — the census equated `hub_indegree` with a specificity test; they are opposite where it matters
+
+**Status:** FINDING (measured, review-driven). Raised as codex P2 on #1079,
+re-measured against source before acceptance, and CONFIRMED.
+**Confidence:** High — read off `nars/tactics.rs:191-215` + `112-124` and the
+consumer's admission predicate, not argued from names.
+
+`COGNITIVE-FABRIC-CENSUS-2026-08-30.md` §"consumer taxonomy ↔ Frontier" carried
+a row asserting *"a marker shared by >1 rival discriminates nothing"* ≡
+`hub_indegree` middle-term exclusion, labelled **"the same computation"**. It is
+not, and the disagreement sits exactly in the region a differential turns on:
+
+| shared by | consumer specificity test | `rcr_abduce` candidates | `rcr_abduce` gaps |
+|---|---|---|---|
+| 1 rival | **specific** → ranked | none (`members.len() < 2`) | none for that predicate; one frontier-level `NoSharedMiddle` iff NO predicate reaches 2 and no hub was seen (`:258-264`) |
+| exactly 2 | **not specific** → disqualified | **permitted**, not guaranteed — self-statement skip, `c_min`, `budget` still apply (`:224-239`) | `BudgetExhausted` if the budget is hit |
+| > `hub_indegree` | not specific | none | `HubExcluded` |
+
+(The candidate/gap split is the second review pass; the first version of this
+table wrote "no gap" and "generates the frontier", conflating a per-predicate
+skip with a frontier outcome and a permission with a result. The inversion the
+entry is about survives unchanged.)
+
+`hub_indegree` lives in `Throttle` beside `c_min` and `budget` — a configurable
+**flood control on hubs**, sharing the *shape* of a sharing-count test and
+inverting its *verdict* below the threshold. Second half, also confirmed:
+`ReasoningGap` is `{kind, subject: Option<u16>, predicate: Option<u16>}`, so
+`NoSharedMiddle` names **no concrete missing term** — the row's parenthetical
+overstated what the gap arm carries.
+
+**The generalizable rule: a throttle and a discriminator can agree in the tail
+and disagree everywhere else.** Both reduce a candidate set; only one does it
+*because the evidence failed to separate*. Structural correspondence between
+two surfaces is a hypothesis to measure at the boundary values, never an
+equivalence to assume — and a census that records one as "the same computation"
+hands a later ownership wave a false BUY.
+
+**Companion finding, uncomfortable and recorded rather than quietly fixed:** the
+same census banks `F-ARW-0` ("census by semantics, not by type name") as its
+own §1 finding 5, then one section later claimed the consumer's reasoning path
+"never claims into" the alpha overlay — on the strength of a name-based grep
+over two files. Measured, the step trace claims indirectly through the resident
+ontology walk; only the differential beside it does not. **Writing the rule down
+is not the same as being immune to it**; the narrowed claim is the one that is
+actually true, and it is also the one that is actionable.
+
+---
+
+## 2026-08-30 — E-BELIEF-ARENA-DEDUP-IS-PAYABLE-AND-W0-MUST-CITE-IT-1 — a ruled duplication the ownership census omits is a re-derivation waiting to happen
+
+**Status:** FINDING (measured). Full census:
+`docs/architecture/COGNITIVE-FABRIC-CENSUS-2026-08-30.md` (three-repo,
+`F-ARW-0`-clean; private-side rendition MedCare-rs PR #596).
+**Confidence:** High — counted at file:line across lance-graph + OGAR + the
+consumer repo, not argued.
+
+`TD-DEEPNSM-V2-BELIEF-DUP` (TECH_DEBT.md:752) ruled the V0 arena
+(`deepnsm-v2/src/belief.rs`) superseded by the canonical planner arena
+(`nars/belief.rs`), held open on two blockers. Measured today: **blocker (b)
+is clear** — external consumers of the V0 re-exports across all three repos =
+ZERO (sole reference: `deepnsm-v2/src/lib.rs:59`); and **the copies have
+diverged** — `admit_derived` exists only in the planner arena
+(`nars/belief.rs:226`), so the dup now compounds with every planner-side
+extension. Yet #1078's W0 ownership-census list names neither
+`nars/belief.rs` nor the TD id (verified against both open plan heads) —
+while its §7 lists NARS revision among loco atom candidates, and an atom
+needs exactly ONE backing arena. Consequence: W0 imports the TD as a
+`SOURCE FACT` row; the dedup (the TD's own payment path: deepnsm-v2 emits the
+stream, consumes the planner arena, deletes `belief.rs` + re-exports, citing
+the D-DIA-V0 tests) precedes any atom freeze. Companion lesson banked in the
+census §7.5: a name-based single-repo grep produced a false absence claim
+about the consumer's same-address overlay within the same session — census by
+semantics, not by type name.
+
 ## 2026-08-27 — E-THE-FUSED-PAYLOAD-IS-INERT-AT-EVERY-EXECUTION-GATE-THAT-CONSUMES-IT-1 — a required field no consumer reads does not stay empty, it fills with fiction
 
 **Status:** FINDING (measured). Deliverable D-MCAL-1. Full entry:

.claude/v3/knowledge/persona-vs-rung-ladder.md

@@ -7,6 +7,30 @@
 > transcript); board anchor: `E-RUNG-CONTENT-LADDER-1` in
 > `.claude/board/EPIPHANIES.md`.
 
+> **⊘ SCOPE CORRECTION (operator, 2026-08-30 — styles are NOT confined to
+> rung 4).** Sessions inheriting this file began reading row 4 as "thinking
+> styles live at rung 4 ONLY." The operator: *"'Thinking styles: rung 4
+> only' is a stale inheritance … it was never a recent ruling."* What this
+> file rules is the CONTENT taxonomy — which kind of thinking OBJECT
+> occupies each rung of the ladder — and that stands unchanged. What it
+> does NOT rule, and never did, is a confinement of style APPLICATION to
+> one rung. The ruled direction (same date): **(a)** rung 0–9 awareness is
+> **to be** scheduled in parallel as meta-aware thinking orchestration
+> through `lance-graph-supervisor` — `kanban_actor.rs`'s `PhaseCensus` (the
+> read-only fleet visibility surface, one `&self` pass) is the built
+> primitive such a scheduler would compose OVER, for the rung above watching
+> the kanban of the layer below. ⚠ The rung-aware dispatch LOOP itself is
+> NOT shipped (`PhaseCensus` and kanban are deliberately rung-blind;
+> `probe_parallel_rung` establishes representation coexistence only, "does
+> NOT establish wall-clock or thread parallelism") — this is a direction,
+> never an available scheduler;
+> **(b)** thinking styles CONVERGE onto the execution membrane — mask-ALU
+> cheapness at the first level, `ogar-loco`/`ogar-r2il` composition for
+> depth — a style is a program over the atom/mask substrate, applicable at
+> any rung. O2 (style → recipe selection) survives, generalized to run from
+> any rung's orchestration into the recipe layer. Census anchor:
+> `docs/architecture/COGNITIVE-FABRIC-CENSUS-2026-08-30.md` §5.5.
+
 ## Why this file exists
 
 The workspace carried FOUR similar-looking spaces that sessions kept
@@ -64,7 +88,7 @@ its "runbook" naming and the board label cross storylines.
 
 ## The four spaces, one line each (anti-conflation card)
 
-1. `StyleFamily` (12) — rung-4 thinking macros, orchestration dispatch. Canonical, deduped, frozen ordinals (driver order, Deliberate=0).
+1. `StyleFamily` (12) — rung-4 thinking macros **by CONTENT taxonomy** (which object occupies the rung), orchestration dispatch. Canonical, deduped, frozen ordinals (driver order, Deliberate=0). ⊘ Not a confinement of style APPLICATION to rung 4 — see the scope correction at the head of this file, and O9 on the address vs the vocabulary.
 2. `recipes::RECIPES` (34) — rung-3 NARS reasoning tactics. THE runbooks. Catalogue in contract; primitives in ndarray `hpc/styles`.
 3. 144 verbs (12×12) — rung-2 thinking atoms, the unified predicate surface (`sigma_rosetta` + `verb_table`). Computing, test-pinned.
 4. persona-36 adjectives — persona-modeling storyline. Off-ladder, unwired, future work. Keep out of rung vocabulary.
@@ -75,12 +99,21 @@ its "runbook" naming and the board label cross storylines.
   not content occupants. No rung references `sigma_rosetta`,
   `verb_table`, `recipes`, or `StyleFamily`. The content ladder exists
   only as this ruling.
-- **O2 — the rung-4→rung-3 edge (`StyleFamily` → recipe selection)
-  does not exist.** A macro choosing which of the 34 tactics fire is
-  the true "default_runbook" semantics; today's `default_runbook()`
-  points at the persona vocabulary instead. Eventual fix: retarget (or
-  twin) the mapping at `recipes::Recipe`, and rename the persona bridge
-  honestly (e.g. `default_persona()`).
+- **O2 — the orchestration→recipe edge (`StyleFamily` → recipe
+  selection) does not exist.** A macro choosing which of the 34 tactics
+  fire is the true "default_runbook" semantics; today's
+  `default_runbook()` points at the persona vocabulary instead.
+  Eventual fix: retarget (or twin) the mapping at `recipes::Recipe`,
+  and rename the persona bridge honestly (e.g. `default_persona()`).
+  ⊘ **Re-scoped 2026-08-30 to match the scope correction at the head of
+  this file** (review finding on #1079, valid): this was written as a
+  "rung-4→rung-3 edge" with a rung-4-specific fix, which contradicts the
+  correction that style APPLICATION is not confined to one rung. The
+  missing edge is **any-rung orchestration → `recipes::Recipe`** — a
+  stratum at any rung selects the tactics it runs. What does NOT change:
+  the edge is still absent, and `default_runbook()` still points at the
+  persona vocabulary rather than the recipe layer. Only the address of
+  the source end generalizes; the defect and the fix are the same.
 - **O3 — persona storyline unwired by design.** Future persona work
   picks up the adjective-36 + `thinking-engine/persona.rs`; until then
   no rung/dispatch code should consume it as if it were reasoning.