D-PUZZLE-0 step 3: crossword propagation as Cartesian-addressed masking, four arms, cui-bono ablation

.claude/board/LATEST_STATE.md

@@ -1,3 +1,8 @@
+## 2026-10-07 — lance-graph #1382 and #1384 MERGED; D-PUZZLE-0 step 3 on branch
+
+- #1382 (`bccc46e`): Sudoku population-fold probe. #1384 (`23e8d93`): shared fold module + crossword probes (steps 2/2b). No contract change in either.
+- Step 3 (branch `claude/epistemicstate5-ce64-bits-3oe0dq`, unmerged): `crossword_mask_propagation_probe.rs`, four representations of one fold; `deepnsm-v2` is now a dev-dependency of `cognitive-shader-driver`. No contract type added: `MooreSymbol8` is probe-local.
+
 ## 2026-10-07 — lance-graph #1381 MERGED (`7334ae9`)
 
 The D-EPI-CANON-0 / D-EPI-MIG-0 section below is now on `main`, together with

.claude/board/PR_ARC_INVENTORY.md

@@ -1,3 +1,16 @@
+## 2026-10-07 — lance-graph PR #1384 (merged `23e8d93`, branch `claude/epistemicstate5-ce64-bits-3oe0dq`, head `94499c3`, 9 commits) — D-PUZZLE-0 steps 2/2b: crosswords on the Sudoku population fold
+
+- **Added:** `examples/shared/population_fold.rs` (the domain-free half: propagation reading, five `facts_population` questions, three folds, declaration gate; a fence test bans domain words); `crossword_population_fold_probe.rs` (synthetic 3-letter alphabet); `crossword_real_words_probe.rs` (DeepNSM v1 4096 + `academic_20k`, DeReKo top 20k from `DEREKO_PATH`; NYT-rule grids; Down-and-Across answers from `XWORD_ANSWERS_DIR`).
+- **Locked (test-pinned):** Sudoku counts unchanged after the refactor (1,000,066 edges); every crossword count equal three ways; disable runs red.
+- **Deferred / open:** the 2b NYT-grid and crossword-answer modes (`94499c3`) were never run end to end; step 3 supersedes the grid generator.
+- **Confidence:** high for the shared-fold claim; the 2b real-word numbers are one run each.
+
+## 2026-10-07 — lance-graph PR #1382 (merged `bccc46e`, branch `claude/epistemicstate5-ce64-bits-3oe0dq`, head `55f2d0f`, 3 commits) — D-PUZZLE-0 step 1: Sudoku at scale on the EpistemicState5 population fold
+
+- **Added:** `sudoku_population_fold_probe.rs`: ~1M claim edges in the propagation reading (GIVEN 20 / FORCED 21 / ENTAILED 22 / CANDIDATE 4); five questions as `facts_population` masks, counted by filter, histogram and per-edge decode.
+- **Locked (test-pinned):** all counts equal three ways and to the generator's counters; the probe is a registered test target (Codex P2).
+- **Confidence:** high; timings one machine.
+
 ## 2026-10-07 — lance-graph PR #1381 (merged `7334ae9`, branch `claude/epistemicstate5-ce64-bits-3oe0dq`, head `0ba5d24`, 17 commits) — CE64 bits 59..63 = `EpistemicState5 = Topology2 × Certification3`; facts populations; legacy deprojection
 
 - **Added:** `lance_graph_contract::epistemic_state5` — `Topology2`, `Certification3`, per-factor facts, `EpistemicState5` (`raw5 = topology | certification << 2`, 24 meaningful / 8 reserved), `Epi5Reading`, `Epi5Declarations::project_state5`, `facts_population` / `Population` / `bit()`; `ClassView::epistemic_reading`; `CausalEdge64::{epistemic_raw5, with_epistemic_raw5}`, `EPISTEMIC_MASK`.

.claude/board/STATUS_BOARD.md

@@ -48,7 +48,7 @@ Plan: `.claude/plans/2026-10-06-global-sudoku-replayable-orchestration-v1.md`. O
 | **D-EPI-CANON-0** | Canonical joint `EpistemicState5` reading of bits 59..63 in the contract | Shipped (#1381) | `contract::epistemic_state5`: Cartesian `Topology2 × Certification3`, `raw5 = t \| c << 2`, 24 meaningful / 8 reserved; per-factor facts; declarations + `project_state5`; `CausalEdge64::{epistemic_raw5, with_epistemic_raw5}`; non-declared-reading writers deprecated |
 | **D-EPI-POP-0** | Facts → 32-state population: the missing operator between one edge's coordinate (u5) and a semantic condition over all coordinates (u32) | Shipped (#1381) | `contract::epistemic_state5::{Population, facts_population, EpistemicState5::bit}`: `const fn`, no DSL / query object / planner. Tests pin exact sets (`CAUSES` = {20..23}, `IND_UNKNOWN \| RELATED` = {10,14,18,22}, `IND_KNOWN \| SUPPORTS` = {13,17,21}, `RELATED` = 8..23), operator == state enumeration over all 2048 requirements with no reserved code, conjunction == intersection; 2 disable runs red (any-of semantics; reserved read as fact-less) |
 | **D-EPI-LEGACY-DEPROJECT-0** | Deproject the remaining `ReasoningBand` / CE64-`TrustTexture` consumers: each goes to an exact coordinate, a `facts_population(..)`, MUL, or another subsystem (e.g. counterfactual context) — never a 1:1 replacement enum | Shipped (#1381) | `entries/2026-10-07-epi-legacy-deprojection.md`: 6 ReasoningBand caller sites (1 certification question, 1 non-interference witness → both exact coordinates; 2 context/obsolete → compat; 2 refusal witnesses), 0 production CE64 `TrustTexture` consumers, 4 `TrustTexture` types with no bridge; `EdgeRole::band` removed; hinge + V3 lift migrated; 4 code loops → `facts_population` |
-| **D-PUZZLE-0** | Sudoku / crossword / chess on the same population algebra under domain laws | In progress | step 1 Sudoku shipped (#1382). Step 2 crossword (`crossword_population_fold_probe.rs`, `entries/2026-10-07-puzzle-0-crossword-same-algebra.md`): domain-free half moved to `examples/shared/population_fold.rs` (fence test, Sudoku counts unchanged); 1,000,052 crossword edges, counts equal 3 ways; filter 0.30–0.43 ns/edge in both domains. Step 2b real words (`crossword_real_words_probe.rs`, `entries/2026-10-07-puzzle-0-crossword-real-words.md`): English = DeepNSM v1 4096 + academic_20k (committed), German = DeReKo top 20k only from `DEREKO_PATH` (CC BY-NC, nothing committed); ~1.0M edges each, counts equal 3 ways. Next: chess in `stockfish-rs` (GPL) — how it reaches the shared half is open |
+| **D-PUZZLE-0** | Sudoku / crossword / chess on the same population algebra under domain laws | In progress | step 1 Sudoku shipped (#1382). Step 2 crossword (`crossword_population_fold_probe.rs`, `entries/2026-10-07-puzzle-0-crossword-same-algebra.md`): domain-free half moved to `examples/shared/population_fold.rs` (fence test, Sudoku counts unchanged); 1,000,052 crossword edges, counts equal 3 ways; filter 0.30–0.43 ns/edge in both domains. Step 2b real words (`crossword_real_words_probe.rs`, `entries/2026-10-07-puzzle-0-crossword-real-words.md`): English = DeepNSM v1 4096 + academic_20k (committed), German = DeReKo top 20k only from `DEREKO_PATH` (CC BY-NC, nothing committed); ~1.0M edges each, counts equal 3 ways. Steps 2/2b merged (#1384). Step 3 (`crossword_mask_propagation_probe.rs`, `entries/2026-10-07-puzzle-0-crossword-mask-propagation.md`): Morton cell + `(slot:offset)` lanes, DeepNSM-v2 `WordId` masks, `MooreSymbol8` letters; four arms on identical puzzles (EN 5×5: token 3.0 µs ≈ hybrid 2.9, Cartesian 5.1, literal ~1,045); with no repeated words, creation stops at 5×5 with the 18–20k vocabularies; cui-bono ablation of givens. Next: chess in `stockfish-rs` (GPL) — how it reaches the shared half is open |
 | **D-EPI-MIG-0** | Migrate readers/writers to the canonical reading; #1378 → conformance | Shipped (#1381) | `entries/2026-10-07-epi-mig-0.md`: affordance law (per-factor compile) + P7a (identity packing) + 6 probes + dismech + V3 lift; conformance census 24/0/8 (was 3/5/16); 3 historical-band probes cannot migrate without invention |
 | **D-EPI-CONFLICT-0** | (superseded by D-EPI-MIG-0: probe converted to conformance) Pin the disagreement between the D-ACR-7-declared P7a band and the affordance law on CE64 bits 59..63 | Superseded | **Current:** the same probe is the conformance test, census 24 agree / 0 disagree / 8 reserved under the Cartesian V1 (D-EPI-MIG-0). **HISTORICAL — the pre-Cartesian pins below no longer describe the code:** `cognitive-shader-driver/examples/epistemic_reading_conflict_probe.rs`: contract-projected `Causes` reads as affordance code 20 `Related` (topology 0) or refuses (1..3), `COUNTERFACTUAL_PROBE` never eligible; `CAUSES` codes 1 and 17 sit on P7a `Open` / `CausalCandidate`; codes 25, 30 on P7a reserved bands; census over 24 cells = 3 agree, 5 disagree, 16 refuse; the affordance law measures where the contract refuses (undeclared / band `Absent`). Both readings run from shared code (`shared/affordance_law.rs`, and P7a's reading moved unchanged into `shared/certification_reading.rs`, used by `relational_certification_probe` too), never copied. Pins of current state, to become agreement tests when the issue closes (`ISS-EPISTEMIC-READINGS-DISAGREE-ON-BITS-61-63`); 5 tests, 6 disable runs red (3 on the affordance codebook, 3 on P7a's: `Contributes` entailment, a reserved code admitted, `Causes` recoded) |
 | **D-MBX-CUTOVER-0** | ShaderDriver on MailboxSoA without BindSpace | In PR | `entries/2026-10-06-shaderdriver-bindspace-cutover.md`: every ShaderDriver → BindSpace edge classified (production state / shared resource / fallback / test oracle / fossil / dead). In `run()` the only mailbox-mode BindSpace read was the ontology; outside it, metrics and the builder read the singleton and `backing()` silently fell back for any mailbox not under id 0. Cutover: optional BindSpace, explicit mailbox selection with no fallback, ontology on the driver, metrics from the selected substrate; no temporal change. `tests/mailbox_cutover.rs` 8 tests, 8 disable runs red |

.claude/board/entries/2026-10-07-puzzle-0-crossword-mask-propagation.md

@@ -0,0 +1,140 @@
+# 2026-10-07 — D-PUZZLE-0-XWORD-MASK (D-PUZZLE-0 step 3): crossword propagation as Cartesian-addressed masking
+
+**Probe:** `cognitive-shader-driver/examples/crossword_mask_propagation_probe.rs`,
+over `examples/shared/population_fold.rs` (unchanged).
+**Status:** MEASURED (one machine, release, `avx2=true avx512f=false`),
+TEST-PINNED (20 tests; 13 disable runs, listed below, all red).
+
+## Question
+
+Once a crossword is compiled, is propagation the Sudoku operation
+(`mask → intersect → popcount → promote → expose → adjacent masks → repeat`),
+with nothing crossword-shaped on the hot path? And which physical
+representation of that one logic is best, on identical puzzles?
+
+## Inventory (read-only, before any code)
+
+| need | existing facility | verdict |
+|---|---|---|
+| cell (WHERE) | `Morton8x8::from_xy(col,row)`, u16 | used |
+| slot position | `FacetTier{hi: slot, lo: offset}.as_u16()` | used |
+| crossing | no stored relation (`morton8x8.rs`: "no stored edge"); equality of cell codes, compiled once into `cross[slot:offset]` and `occupant[cell]` lanes | used |
+| word (WHAT) | DeepNSM-v2 `PaletteVocab::from_frequency_ranked`, `WordId = u16` | used |
+| masks | `ndarray::simd::{mask_and_assign, popcount_batch_u64}` (already a dependency) | used |
+| `MaskOp::Gather` (#1383) | foreign-key semijoin | not needed: elimination is an AND within one WordId space |
+| letter | no canonical Moore symbol space: every Moore in the repo is a direction table, a grid-edge validity byte or a palette-law operand | declared probe-local `MooreSymbol8` |
+
+No new primitive and no contract change. `deepnsm-v2` became a dev-dependency
+of the driver (its `guid-v3-tail` is already a default contract feature).
+
+## Encodings
+
+- `MooreSymbol8` (u8): `0` unknown, `1..=26` a–z, `27` ä, `28` ö, `29` ü,
+  `30` ß, `31..=255` reserved. Accented Latin letters fold to the base letter
+  (`cliché` is spelled `cliche`); ä ö ü ß are letters of their own. A word with
+  any other character keeps its `WordId` and is not a crossword word.
+  Promoting this reading to a canonical value tenant is a contract decision,
+  not taken here.
+- English vocabulary: `academic_20k.csv`, lowercased, through
+  `from_frequency_ranked`, exactly as `deepnsm-v2/examples/genre_shapes.rs`
+  builds it: **18,555** ids (`vocab.rs` quotes 18,559 distinct surfaces; four
+  collapse when lowercased), 17,548 crossword words.
+- German vocabulary: no German vocabulary is committed and DeepNSM-v2 builds
+  none. With `DEREKO_PATH`, the 20,000 most frequent DeReKo-2014 forms
+  (lowercased, `NE` and non-alphabetic forms dropped, frequencies summed),
+  19,835 crossword words. CC BY-NC 3.0; counts only here, nothing derived
+  committed.
+- Candidates: `P(L, i, x)` bitsets over a language's `WordId`s; a slot starts
+  at `P_all(L)`. 16.4 MB (EN) / 17.7 MB (DE) of populations for L = 3..=21.
+- State: CE64 bits 59..63 only, the shared GIVEN 20 / FORCED 21 / ENTAILED 22
+  / CANDIDATE 4. Slot and word sit in separate content lanes.
+
+## The four arms (same puzzles, same givens; equal fixed points asserted)
+
+| arm | content | addressing | propagation |
+|---|---|---|---|
+| A token | `WordId` masks | `cross` | letter at offset `i` ANDs `P(L,j,x)` into the crossing slot |
+| B Cartesian | `MooreSymbol8` board | `occupant` | symbols written; affected slots recompute from all their cells |
+| C literal | strings | cell array | every unplaced slot re-scans vocabulary strings each round (steps 2/2b) |
+| D hybrid | `WordId` masks + board | `occupant` | each new symbol ANDs one `P` into the cell's other slot |
+
+## Measured
+
+Final run at `2bdb50b` (both review fixes in). One machine, release,
+`avx2=true avx512f=false`.
+
+Filter, 2000 patterns (length + 1–3 revealed symbols):
+
+| | EN | DE |
+|---|---|---|
+| mask AND chain + popcount | 775 ns/pattern, 11.2 ns/claim | 543 ns/pattern, 6.0 ns/claim |
+| direct spell-lane scan | 35,755 ns/pattern | 42,124 ns/pattern |
+
+Solve, median over puzzles of each arm's best of 5 warm runs, rotated order,
+one warm-up per arm; the literal arm's word list is built once per language,
+outside the timing (µs):
+
+| | A token | D hybrid | B Cartesian | C literal | ANDs A / B | bytes A / B |
+|---|---|---|---|---|---|---|
+| EN 5×5 (30) | 3.0 | 2.9 | 5.1 | 1,045 | 11.3 / 20.4 | 23,220 / 2,381 |
+| DE 5×5 (30) | 2.3 | 2.3 | 3.7 | 986 | 7.6 / 11.4 | 25,060 / 2,568 |
+
+- A and D cost the same: routing letters through the Morton board adds no
+  measurable time. B does about twice the ANDs (it recomputes) and is ~1.6×
+  slower, but holds ~10× fewer bytes (one board instead of a full WordId
+  mask per slot). C, the string method, is ~350–430× slower than A.
+- Two earlier timing methods were wrong and are superseded by the table
+  above: timing each arm once, cold, A first (A showed 16.6 µs against D's
+  6.9 µs at an equal AND count: cache warm-up), and timing the literal arm
+  once including the rebuild of its word list (~5,200 µs; Codex review).
+- The shared fold on a 5×5 lane (~1.0M claims, EN and DE): every count equal
+  three ways; filter 0.34–0.35 ns/edge, histogram ~2.0, decode oracle ~3.6 —
+  the same costs as Sudoku and steps 2/2b.
+
+Creation (budget 100,000 tried words per step, 30 s per size), NYT-rule
+grids, **no word used twice in one fill**:
+
+| side | EN made / misses | DE made / misses |
+|---|---|---|
+| 5 | 30 / 1 | 30 / 4 |
+| 7–21 | 0 (all misses are fill, none uniqueness) | 0 (same) |
+
+The no-repeat rule (CodeRabbit review) changed the 7×7 result. Before it,
+EN made 5 / 175 and DE 2 / 163 boards at 7×7; with it, none. Those fills
+reused a word across slots, which a crossword does not allow. So with the
+18–20k general vocabularies, creation stops at 5×5. Solving is not the limit
+at any size measured; creation is. The likely cause is vocabulary density
+(crosswords are built from answer lists with many short fill words); not
+measured here.
+
+Cui bono, each given removed in turn (fixed point re-run + uniqueness count),
+5×5 only (no larger board was created):
+
+| | givens | Redundant | ShortcutOnly | SilentlyNecessary | Necessary | out of budget |
+|---|---|---|---|---|---|---|
+| EN 5×5 | 116 | 28 | 40 | 0 | 48 | 0 |
+| DE 5×5 | 74 | 14 | 25 | 0 | 35 | 0 |
+
+Mean slots lost from the fixed point per removed given: EN 1.26, DE 0.20.
+`SilentlyNecessary` (a given propagation never used but search needs) did not
+occur in the 190 ablations. The readout nominates; it does not prove cause: a
+given necessary in one set can be replaced by another set.
+
+## Falsifiers (disable-verified red)
+
+Arm A ignores crossings (4 tests); letters a/b swapped (codebook, decode);
+index drops offset 0 (5, incl. index-vs-scan); no promotion at popcount 1 (2);
+first crossing only (every-crossing test); zero candidates not a contradiction;
+language gate removed; word bits written into the edge (CE64 test); crossings
+rediscovered from the cell lane (compiled-lanes test); ablation ignores
+uniqueness, and ablation removes nothing (both ablation tests); no-repeat
+rule removed (fills-never-repeat test).
+
+## Open
+
+- Creation beyond 5×5 with these vocabularies (try a crossword answer list
+  at runtime).
+- Whether `MooreSymbol8` (and a `MooreLearn8` companion) become canonical
+  value tenants: a contract decision.
+- The step-2b NYT-grid + crossword-answer mode (#1384, `94499c3`) was never run
+  end to end; this probe supersedes its grid generator.

.claude/board/entries/README.md

@@ -25,7 +25,7 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-252 entries, 2026-08-06 .. 2026-10-07.
+253 entries, 2026-08-06 .. 2026-10-07.
 
 | date | entry id | finding | file |
 |---|---|---|---|
@@ -34,6 +34,7 @@ exactly one of them, never both.
 | 2026-10-07 | `D-PUZZLE-0` |  | [2026-10-07-puzzle-0-sudoku-population-fold.md](2026-10-07-puzzle-0-sudoku-population-fold.md) |
 | 2026-10-07 | `D-PUZZLE-0-XWORD` |  | [2026-10-07-puzzle-0-crossword-same-algebra.md](2026-10-07-puzzle-0-crossword-same-algebra.md) |
 | 2026-10-07 | `D-PUZZLE-0-XWORD-REAL` |  | [2026-10-07-puzzle-0-crossword-real-words.md](2026-10-07-puzzle-0-crossword-real-words.md) |
+| 2026-10-07 | `D-PUZZLE-0-XWORD-MASK` |  | [2026-10-07-puzzle-0-crossword-mask-propagation.md](2026-10-07-puzzle-0-crossword-mask-propagation.md) |
 | 2026-10-07 | `gated-gather-bit-schedule` | Production gather is 3.8× slow from a per-row branch; bit-gated visitation wins or ties at every density; ordinal vectors never win (R1 holds) | [2026-10-07-gated-gather-bit-schedule.md](2026-10-07-gated-gather-bit-schedule.md) |
 | 2026-10-07 | `D-EPI-MIG-INVENTORY` |  | [2026-10-07-epi-mig-inventory.md](2026-10-07-epi-mig-inventory.md) |
 | 2026-10-07 | `D-EPI-CANON-0` |  | [2026-10-07-epi-mig-0.md](2026-10-07-epi-mig-0.md) |

crates/cognitive-shader-driver/Cargo.toml

@@ -163,6 +163,10 @@ test = true
 name = "crossword_real_words_probe"
 test = true
 
+[[example]]
+name = "crossword_mask_propagation_probe"
+test = true
+
 [dependencies]
 lance-graph-contract = { path = "../lance-graph-contract" }
 # Read-only ontology registry (Phase 7, v4 plan). default-features = false so the
@@ -233,3 +237,6 @@ lab = ["serve", "grpc", "with-engine", "with-planner"]
 # D-CTX-0: the one Fisher-Z calibrator, for the probe only. Production code
 # borrows the calibrated bytes; it never depends on the calibrator.
 bgz-tensor = { path = "../bgz-tensor" }
+# D-PUZZLE-0 step 3: crossword candidates are DeepNSM-v2 `WordId`s, assigned by
+# its own `PaletteVocab::from_frequency_ranked`. Probe-only.
+deepnsm-v2 = { path = "../deepnsm-v2" }

crates/cognitive-shader-driver/examples/README.md

@@ -442,3 +442,19 @@ edges per language; every count agrees three ways. Board:
 cargo run --release -p cognitive-shader-driver --example crossword_real_words_probe
 DEREKO_PATH=/path/DeReKo-2014-II-MainArchive-STT.100000.freq cargo run --release -p cognitive-shader-driver --example crossword_real_words_probe
 ```
+
+## crossword_mask_propagation_probe.rs
+
+D-PUZZLE-0 step 3. Is crossword propagation the Sudoku operation, and which
+representation of it is best? Cells are `Morton8x8` codes, slot positions
+`(slot:offset)` tiles, crossings two compiled u16 lanes, letters a declared
+`MooreSymbol8` byte, candidates DeepNSM-v2 `WordId` bitsets. Four arms (token,
+Cartesian board, literal strings, hybrid) solve the same created puzzles to the
+same fixed point; the run also ablates each given (cui bono) and folds a ~1M
+claim lane the shared three ways. German needs `DEREKO_PATH`. Board:
+`.claude/board/entries/2026-10-07-puzzle-0-crossword-mask-propagation.md`.
+
+```
+cargo run --release -p cognitive-shader-driver --example crossword_mask_propagation_probe
+```
+