Skip to content

Synergy map after #378–#383; guard symlink fix; stale mul_add doc #384

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AdaWorldAPI merged 2 commits into from
Oct 29, 2026
+270 −4
Merged

Synergy map after #378–#383; guard symlink fix; stale mul_add doc #384

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c75a67b
guard: skip dangling symlinks; correct stale wasm mul_add and DR512 n…
claude Oct 29, 2026
de61987
knowledge: synergy map 2026-10-29, revising the 10-16 listing after #…
claude Oct 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion .claude/hooks/facet-stride-guard.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,7 +217,9 @@ print(ti.get("file_path") or ti.get("notebook_path") or "")' 2>/dev/null)"
for f in "$@"; do f="${f#"$ROOT"/}"; [ -f "$f" ] && whole_file "$f" | check_lines "$f"; done
;;
--all)
while IFS= read -r f; do whole_file "$f" | check_lines "$f"; done \
# A tracked path can be a dangling symlink (crates/burn/src/*.rs point
# into an absent ../upstream checkout); skip anything that is not a file.
while IFS= read -r f; do [ -f "$f" ] && whole_file "$f" | check_lines "$f"; done \
< <(git ls-files '*.rs' | while IFS= read -r f; do in_scope "$f" && echo "$f"; done)
;;
*) echo "usage: $0 --hook | --diff <rev> | --all | --file <path>..." >&2; exit 64 ;;
Expand Down
250 changes: 250 additions & 0 deletions .claude/knowledge/synergy-map-2026-10-29.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,250 @@
# Synergy map 2026-10-29 — since lance-graph #1591 and ndarray #378

**Status:** OPEN — exploration map. It decides nothing; the operator chooses.

**Base:** lance-graph `.claude/board/entries/2026-10-16-synergy-map-since-1579.md`
(the last detailed mapping). That entry revised the 2026-10-11 fleet listing
(`coresearch-2026-10-11-simd-fleet.md`: A–F, X1–X19). This map revises every
point of it and adds what the merges since then closed or opened.

**Delta read:**
- ndarray: #378–#383, all merged (`cadb572` … `2cbe131`), plus this PR.
- lance-graph `main` `b272a1c`: 114 merges since 2026-10-16 18:00. Only the
rows this listing tracks were re-read.
- Fleet default branches, fetched 2026-10-29: lichess-rs `79af1ab`,
tesseract-rs `7cbc969`, lance-graph-java `d8cdcc1`, OGAR `2a080a9`,
MedCare-rs `fb829b6`, q2 `66281e7`, ladybug-rs `7f2cc38`, surrealdb
`ef75a05`.

**Method:** Haiku cards did the harvests (§7). Every row marked CLOSED or
LOOSE below rests on a line that was re-read on the main thread; rows marked
"not re-read" are carried over unchanged.

## 0. What the six ndarray PRs did

| PR | closed | opened or left |
|---|---|---|
| #378 | wasm `F32x16`/`F64x8::mul_add` round once: certified fast domain, local repair, exact fallback. `+relaxed-simd` tested and identical. Proof obligations in `wasm-fma-proof-obligations.md` | `reduce_sum` order still differs per realization; no mechanised proof |
| #379 | NaN-exclusion mindmap and premise audit (document) | S2 masked float fold, S3 `n_excluded` |
| #380 | Palette512 / SM9 scalar probe; note corrected (tests run, VPDPBUUD exists) | its separate SIMD gate |
| #381 | float lanes: presence IS NaN, held as a bijection (`not_nan_*_to_mask`, `nan_where_clear_*`, `presence_nan_mismatches_*`); SQL TRUE planes `{lt,…,ne}_{f64,f32}_to_mask`; `mask_not_under`; SM9 probe with the 3σ gate | the two-plane SM9 layout (superseded by #383) |
| #382 | FACET STRIDE rule, guard and exclusions; `simd_stride::{lut_u8, lut9}` | burn-down list: 166 ndarray hits, 10 lance-graph hits |
| #383 | DR512 P×M law: ONE presence/polarity bit (P=0,M=0 is the only NaN); `hpc::codec9` (`EpiLane9`, `Codec9::encode_touched`); `lut9` moved onto one plane; CI job `dr512/epistemic-presence` runs the 512-codeword, distance, WHERE-64K and ceiling oracles plus our probes | CausalEdge64 512-palette sidecar DEFERRED; LUT distance speed unmeasured |
| this PR | guard `--all` skipped nothing on dangling symlinks (34 `crates/burn/src/*.rs` → absent `../upstream`), now skips them (count unchanged at 166); `wasm-simd-consumer-guide.md:206` said wasm `mul_add` rounds twice; mindmap records #383 | — |

## 1. The base listing, point by point

### A. `fearless_simd` vs `ndarray::simd`

| point | 10-16 | now | evidence |
|---|---|---|---|
| compile-time dispatch, no per-kernel runtime dispatch, v4-on-AVX2 SIGILL, 512-shaped names, SSE tiers, AVX2 integer polyfills | CLOSED | CLOSED | no change |
| aarch64 guard | LOOSE | LOOSE | `cpu_guard.rs:318` still says it needs `getauxval(AT_HWCAP)`; the only `getauxval` call reads `AT_PHDR` (`:461-465`) |
| generated backends | PARTIAL | PARTIAL | — |
| per-level parity, 18 types | LOOSE | LOOSE, refined | of the 18, 16 never occur in `crates/simd-masking-parity/src/lib.rs`; `I16x32` occurs 7 times, `I8x64` once (counted, not judged as coverage) |
| MSRV prose | LOOSE | LOOSE | lance-graph `ISSUES.md:472` `ISS-NDARRAY-CANONICAL-COORDINATE-COUPLES-FLEET-MSRV`: **Status: OPEN** |
| one binary everywhere | EXCLUDED | EXCLUDED | — |

### B. "Three places the CPU leaks back"

| point | now | evidence |
|---|---|---|
| the build cares; runtime dispatch | CLOSED | — |
| consumers inherit no config | CLOSED for lance-graph; LOOSE for lance-graph-java | lgj `native/lgj-abi/.cargo/config.toml:35` still says "one known host"; `check_build_cpu` has 0 hits on lgj `main`. The v4 hard-code was not re-read |

### C. Accelerate split — PARK, unchanged. ### E/F (SSE slow paths) — unchanged.

### D. Unsafe inventory

Unchanged: missing `// SAFETY:` comments, CONSOLIDATE / SAFE-API rows, the
SAFE-CRATE (`bytemuck`/`zerocopy`) question. `is_x86_feature_detected!`:
96 sites in 17 files (counted today), still unclassified into
startup / once / per-call.

### Findings X1–X19 and TD rows

| id | 10-16 | now | evidence |
|---|---|---|---|
| X1, X7, X18 | CLOSED | CLOSED | X18 re-read: OGAR takes `encryption` from ndarray `branch = "master"` (`ogar-encryption/Cargo.toml:15`), the named-branch form the substrate law allows |
| X2–X4 ladybug | LOOSE | **UNKNOWN** | a search on ladybug-rs `main` for `grep avx512f`, `x86-64-v2`, `config-v2` found nothing. Search is not evidence of a fix; read the image selector before closing |
| X6 v4 Railway images | DECIDED | DECIDED | q2 `Dockerfile:227` still `-C target-cpu=x86-64-v4`; MedCare-rs: no `x86-64-v4` string found in Dockerfiles/toml/yml (not read; may have moved) |
| X11 `mul_add` | LOOSE (doc pair contradicted) | **CLOSED** | `simd.rs:1319` "fused on every arm" now matches `simd_wasm.rs:37-40` (emulated, rounds once). The last stale doc, `wasm-simd-consumer-guide.md:206`, is fixed in this PR. lance-graph CP-3 (harvest 10-27) asked for a "parity statement": parity group `check_mul_add_single_rounding` plus the proof document is that statement |
| X16 consumer raw intrinsics | LOOSE | LOOSE, **merged with the FACET STRIDE lance-graph hits** | lance-graph CP-2 (10-27) ADOPTED: `ndarray_bridge.rs` raw blocks sit under `cfg(not(feature = "ndarray-hpc"))` (13 occurrences), a non-default fallback. `TD-SIMD-SWEEP-W2` HALF DONE: `types.rs` migrated, `ndarray_bridge.rs` 66 raw lines left. The guard's 10 lance-graph R1 hits are exactly those blocks (8) plus holograph `hamming.rs:593-594` (2, live) — one burn-down, not two |
| X17 forbidden remedies in records | PARTLY CLOSED | PARTLY CLOSED | still open: lichess-rs `CLAUDE.md:32,60`, `README.md:93-97`, `engine-adapter-stockfish/Cargo.toml:14,16` say `+1.95`; tesseract-rs `rust.yml:60` "ndarray's rust-version = 1.95 floor"; the MSRV issue |
| X19 lgj | LOOSE | LOOSE | see B |
| TD-NDARRAY-SIMD-POPCNT-NATIVE | (b) CLOSED | (b) CLOSED | (a) objdump and (c) the `Cargo.toml` comment not re-read |

## 2. The 10-16 dots D1–D8

| dot | 10-16 | now | evidence |
|---|---|---|---|
| D1 one set algebra, four spellings | LOOSE, hinge D-RA-2 | **CLOSED by decision and code** | D-RA-2 DECISION 2026-10-20 option A; D-CMR-1 shipped (#1644, relalg `normalize` delegates to `contract::set_algebra`); D-CMR-6 shipped (#1654, relalg `symbolic` delegates to `contract::set_rules`). crepe's join-body rewrites stay, by the decision |
| D2 one preimage op | CLOSED AS VIOLATION | CLOSED | — |
| D3 read masks × clustering × prefilter | LOOSE | not re-read | — |
| D4 per-row truth lane terminal | LOOSE | LOOSE | D-NALP-4 "Partial, shipped (#1548)": scalar law exact on 2³² rows; whether a lane terminal is needed is open |
| D5 one `Binder::code` | LOOSE | LOOSE | D-ASM-1b Queued, not authorized |
| D6 SAG-5 | LOOSE | LOOSE | D-SAG-5-0 … -10 Queued / docs |
| D7 two ndarray ids | CLOSED | CLOSED | — |
| D8 doc drift | PARTLY CLOSED | PARTLY CLOSED | closed: X11 pair, consumer guide (this PR). Open: X19 lgj comment, MSRV issue option, lichess/tesseract comments |

## 3. The 10-16 synergies S1–S8

| | 10-16 | now | evidence |
|---|---|---|---|
| S1 one register, five reasoners | LOOSE | LOOSE, **unblocked** | its prerequisite D-CGW-1 shipped (#1594); D-CGW-13 Queued |
| S2 wire gate = admission-as-mask | LOOSE | LOOSE, **new ndarray hook (N8)** | D-RPF-4 "Blocked (4 decisions; T1 LUT-weighted strided sum gap)" |
| S3 three devices, one PCS | LOOSE | LOOSE | `ISS-QUALIA-A2B-CLUT-UNMEASURED` open; D-CGW-14 Queued (modulators only) |
| S4 orientation, two readers | LOOSE | **CLOSED** | #1594: `causal_edge::Orientation` on 43..45; Moore `lift_ce64` refuses an oriented edge, so there is one table |
| S5 Binder is the one bind surface | LOOSE | LOOSE | D-ASM-1b Queued |
| S6 one address minting path | LOOSE | not re-read | — |
| S7 `syllogize` writes 50..52 as `0b111` | LOOSE | LOOSE | the D-CGW-1 row says "46..49 (D-CGW-9) and 50..52 still open"; `syllogize` now lives in `edge_v3.rs` / `isa.rs`; whether 50..52 is still written as `0b111` was not read |
| S8 triad readers to migrate | LOOSE | **CLOSED** | #1594 removed the S/P/O triad ("no production writer") |

## 4. New synergies since then (the dots between arcs)

**N1 — `lut9` and Codec9 have their first consumer, and it adds no kernel.**
lance-graph `planner/src/nars/scent_learn.rs:49-50` imports
`ndarray::hpc::codec9::EpiLane9` and `ndarray::simd::lut9` (D-EPI-COT C3c,
TEST-PINNED). Its learning rate is the `lut9` table; its delta is built in
code space with the same zero rule `lut9` applies to its output. The
dependency direction is the right one: the consumer names ndarray words and
grows none.

**N2 — one P×M rule, three homes, two with a pairing test.** The rule
"absent = P=0 AND M=0; 9-bit P is also the sign; touched zero is +0" lives in
ndarray `codec9.rs`, in lance-graph-contract `lane_codec.rs` (`EpiLaneSpec`,
presence plane mandatory, a 9-bit lane must be Sym), and in the
thought-layers probe (`write_epi`: a NaN lane is not written, merge on read
keeps the prior). The contract tests that it "agrees with `codec9` on all 512
codes" (`lane_codec.rs:993`). The third copy, `lance-graph-nsl/src/eval.rs:3-7`,
mirrors Codec9 with a runtime scale and has no such test (the crate is
dependency-free by design).

**N3 — the float bridge is one mask expression.** A DR512 lane's float
presence is `P | ne_u8_to_mask(M, 0)`; its complement is "never touched".
That is exactly the plane #381's `nan_where_clear_f32` consumes, so decoding a
DR512 lane to f32 keeps #381's bijection with no new primitive. The probe test
`decode_satisfies_the_float_bijection` already pins this end to end.

**N4 — NSL wants a runtime-scale, batch Codec9.** D-NSL-W0W1 kept "SM9 as a
storage format for neighbor input features" as a PROBE (9-bit 2–4× better
than int8) that "needs Codec9 runtime scale, batch encode, and a Cora-scale
run first". ndarray's Codec9 is scalar and fixes the scale at compile time on
purpose (`codec9.rs:18-21`: "Storage, polarity, Sym and NaN policy are
intentionally NOT configurable … the only compile-time knobs are … rounding
mode and scale calibration"). A runtime scale is therefore a ruling, not a
patch. NSL also REJECTED SM9 edge weights as sort keys (top-1 changes for
7–21% of seeds); that verdict stays.

**N5 — the "touched presence" word.** `scent_learn.rs:345-356` builds
`P = present & (positive | (M == 0))` with a scalar loop over its features;
`lut9` applies the same rule internally. With a batch encoder (N4) this would
be the second writer of the same word; the facade could then name it
(`present & (positive | eq_u8_to_mask(M, 0))`, two mask ops). One consumer
today, so not yet.

**N6 — D-RPF-4's "T1 LUT-weighted strided sum gap" may already compose.**
ndarray now has `lut_u8` (facet stride) and `masked_strided_group_sum`. A
LUT-weighted strided sum is `lut_u8` into a weight lane, then the strided
group sum. Whether that meets D-RPF-4's need is unread; a probe answers it
before anyone builds a fused primitive.

**N7 — X16 and the FACET STRIDE burn-down are one job.** The 8 lance-graph
R1 hits are the dead fallback blocks of `TD-SIMD-SWEEP-W2`; deleting or
migrating them (to `U64x8::popcnt` / `ndarray::hpc::bitwise`) closes both
lists at once. holograph `hamming.rs:593-594` is the only live one.

**N8 — the popcount question the rule raises.** All 7 ndarray R1 hits and all
10 lance-graph R1 hits are the nibble-table popcount (`pshufb` over two 4-bit
halves). On AVX-512 VPOPCNTDQ replaces it; on AVX2 there is no vector
popcount, and the nibble table is the published fast method (Muła et al.).
The FACET STRIDE rule forbids a lookup built from shuffles. The options are a
per-u64 scalar `popcnt` (four per 256-bit register), or an exclusion with a
reason. This needs a measurement on an AVX2 host before either is chosen.

**N9 — the NaN erasers stay the open hazard.** `quantize_f32_to_u8`
(`hpc/quantized.rs:517`) and `quantize_f32_to_i8` (`:545`) still map NaN
through a saturating cast; 14 call sites in ndarray, 4 in lance-graph. Under
both rulings (float presence IS NaN, DR512 never-touched) they silently turn
NULL into a value.

## 5. Closed ends, loose ends, fruit

### Closed since 10-16

1. X11 and its doc pair (#374, #378, this PR).
2. D1 (D-RA-2 option A; #1644; #1654).
3. S4 and S8 (#1594, D-CGW-1).
4. H2 superseded: D-CGW-9 replaced by D-HL-1 O1 (46..49 = activation i4).
5. NaN exclusion on float lanes (#381) and integer lanes (#383, one plane).
6. FACET STRIDE rule, guard, audit (#382); guard symlink crash (this PR).
7. `lut9` / Codec9 adopted downstream without a new kernel (N1).
8. Palette512 probe corrected and kept as a CI tripwire (#380, #383).

### Excluded or rejected (do not reopen without new evidence)

- One binary everywhere (operator).
- Runtime dispatch of core types.
- Gather / preimage designs (D2, D-ASM-11/11A/11B, D-QMB-0…5: VIOLATION banners).
- SM9 edge weights as sort keys (D-NSL-W0W1).
- A second DR512 bit interpretation (Asym, Floor, NoNaN) for Codec9 (#383).

### Low-hanging (≤ 1 hour each)

| # | item | repo | check |
|---|---|---|---|
| L1 | lichess-rs `+1.95` (5 places) and tesseract-rs `rust.yml:60` comment → 1.99 | lichess-rs, tesseract-rs | doc only; needs push in those repos |
| L2 | lgj: name the artifact tier, drop "one known host", call `check_build_cpu()` in the handshake | lance-graph-java | needs push approval |
| L3 | NSL: a pairing test between its SM9 mirror and Codec9 semantics (512 codes, the -0.4/-0.5 ceiling cases), kept in NSL's own test data since the crate is dependency-free | lance-graph | the contract's `lane_codec.rs:993` test is the template |
| L4 | restore the facet-stride paragraph in `lut9`'s doc (lost in #383's rewrite) | ndarray | doc only |
| L5 | classify the 96 `is_x86_feature_detected!` sites (startup / once / per-call) as an inventory column | ndarray | grindwork for a Haiku census + Sonnet classification |

### High-hanging (a session each)

| # | item | why | gate |
|---|---|---|---|
| H1 | parity groups for the 16 never-named facade types | A | one flipped scalar op per group goes red |
| H2 | aarch64 guard via `AT_HWCAP` | A | dotprod build under `qemu-aarch64 -cpu cortex-a53` reports, no SIGILL |
| H3 | AVX2 popcount: scalar `popcnt` vs nibble table, measured, then migrate the 7 ndarray R1 hits or exclude them with the number as the reason | N8 | the faster one on a named AVX2 host; parity unchanged |
| H4 | lance-graph `ndarray_bridge.rs` dead blocks deleted, holograph `hamming.rs` onto `ndarray::simd` | X16, N7 | guard reports 0 for lance-graph; holograph Hamming tests unchanged |
| H5 | `quantize_f32_to_{u8,i8}`: refuse NaN or document the mapping; parity test that a NULL=NaN lane survives `Codec8 NaNAtMin` with zero mismatches | N9 | `presence_nan_mismatches == 0` after the round trip |
| H6 | probe: does `lut_u8` + `masked_strided_group_sum` cover D-RPF-4's T1 gap | N6 | equals D-RPF-4's reference fold on its fixture |
| H7 | D-CGW-13 first fixture (ONA + Java OpenNARS on one NAL fixture) | S1, now unblocked | two disagreeing oracles never project to equal edges |
| H8 | S7: read whether `syllogize` still writes 50..52 as `0b111`; if so, give 50..52 its one meaning | S7 | field isolation on 50..52 |

### Heavy-hanging (multi-session, owner or ruling needed)

| # | item | why |
|---|---|---|
| V1 | runtime-scale, batch Codec9 (SIMD encode into magnitudes + presence plane) | N4: a ruling on #383's "only compile-time knobs" |
| V2 | masked float fold with a fixed order, carrier `n, mean, M2` (mindmap S2), plus `n_excluded` (S3) | determinism ladder steps 2–3 |
| V3 | DR512 comparisons without decode (`cmp` on P×M as masks) | waits for a consumer; lance-graph lance13 map X14 parks u8 `Pred` until one asks |
| V4 | CausalEdge64 512-palette sidecar | DEFERRED by #383; lance-graph CE64 owner |
| V5 | per-classid label codebook behind `Binder::code` (D-ASM-1/1b) | D5, not authorized |
| V6 | D-NALP-5…7 | D4 |
| V7 | SAG-5 on D-CMR | D6 |
| V8 | Accelerate split, Kani, `bytemuck`/`zerocopy` | PARK |

## 6. Not searched

ladybug-rs image selector (search empty, not read); MedCare-rs deploy CPU
flags (search empty, not read); lgj `.cargo/config.toml` lines 1–34 and 36+;
S6 minting path; S7 `syllogize` body; D3; the 114 lance-graph merges beyond
the rows above; stockfish-rs, HubSPO-rs, spear, a2ui-rs, surrealdb (fetched,
not searched for these items; surrealdb's `ndarray = "0.17.2"` from crates.io
is unchanged).

## 7. Harvest ledger (Haiku cards)

| card | task | result | spot check on the main thread |
|---|---|---|---|
| substrate-mapper | map of #383's six files | 172 entries, DONE | `codec9.rs` read in full; `:56` `NEVER_TOUCHED`, `:148` `encode_touched` match |
| substrate-mapper | lance-graph consumers of SM9/Codec9 | 151 entries, PARTIAL (STATUS_BOARD too large to read whole; searched) | `scent_learn.rs:49-50` imports and `:345-356` zero rule re-read |
| harvest-scout | NaN / presence / shuffle candidates in lance-graph | DONE, but ran on a stale local checkout (128 commits behind `main`) | its 10 `shuffle_epi8` hits match the guard's 10 lance-graph hits |
| census-counter | guard counts, lance-graph grep counts | DONE; exposed the guard's symlink crash | crash reproduced; fixed in this PR |
| census-counter | X-item status across 9 repos (18 commands) | DONE | `simd.rs:1319`, consumer guide `:206`, `cpu_guard.rs:318` re-read |
| quote-harvester | six lance-graph board entries | DONE, 6/6 | NSL `:20-23` decisions re-read |
| census-counter | board status rows for 20 IDs | DONE | `STATUS_BOARD.md:364,372,494,594`, `ISSUES.md:472`, `TECH_DEBT.md:1947,2274` re-read |

Misses: one (the stale-checkout scout; my brief named the local checkout
before I had fetched `main`). No card miscounted or concluded.
Loading
Loading