docs(knowledge+board): SIMD alien-magic capture — E-SIMD-SWEEP-1 + 13 TD entries + wave plan#400
Merged
Merged
Conversation
…D entries, wave plan
Architectural capture commit per autoattended-multiagent-pattern §3
("P0 fixes land BEFORE the next sprint PR opens"). The simd-savant
PRE-MERGE audit of origin/main surfaced 158 raw-intrinsic violations
across 5 consumer crates (mul.rs, blasgraph/types + bridge, holograph
hamming, bgz17 simd + prefetch, thinking-engine VNNI dispatch) plus
3 missing primitives in ndarray::simd that block clean remediation.
Before any worker spawns to fix PR #398's codex P1/P2 findings, the
architectural shape needs writing so every wave-1 worker briefs
against the same canonical reference.
Files (4):
1. .claude/knowledge/ndarray-vertical-simd-alien-magic.md (NEW, 144L)
The canonical reference. "The Click" statement (ndarray ships
struct methods on typed wrappers + closure-parameterized batch
primitives; consumers compose with domain enums). Per-workload
surface table covering palette L1-L4, spatial splat, blasgraph-
over-palette, i4 packed qualia, hamming over u64, signature
kernels. W1a (5 ndarray PRs) + W1b (5 consumer migrations) +
W1.5 (3 sigker primitives, gated on jc Pillar 11) wave plan.
Cross-links to simd-savant card, autoattended-pattern §14,
sigker lib.rs, Jirak iron rule. Litmus tests for surface
proposals.
2. .claude/board/EPIPHANIES.md (PREPEND E-SIMD-SWEEP-1)
Captures the 158-violation finding: PR #398 was the 5th
violation, not the first. Doctrinal claim: the SIMD source-of-
truth invariant is retroactive, not just forward. Full AP-SIMD-N
breakdown (117 / 8 / 13 / 7 / 19 / 0 / 2 / 13 = 158). Strategic
angle on sigker as Index-regime third lane that bypasses
I-NOISE-FLOOR-JIRAK. Doctrinal counterpart to E-META-10 /
I-LEGACY-API-FEATURE-GATED (same retroactive-sweep shape).
3. .claude/board/TECH_DEBT.md (PREPEND 10 entries)
- W1a (5 entries): TD-NDARRAY-SIMD-UNPACK-I4-16D,
-SATURATING-ABS-I8, -GATHER, -PREFETCH, -POPCOUNT-U64.
Each with severity, required API surface (Required ndarray
PR), and cross-refs.
- W1.5 (3 entries, DEFERRED): TD-NDARRAY-SIMD-SIGNATURE-PDE-SWEEP,
-RANDOMIZED-PROJECTION, -LYNDON-PACK. Gated on sigker
benchmarking + jc Pillar 11 certification.
- W1b consumer migrations (5 entries): TD-SIMD-SWEEP-W1
(holograph), W2 (blasgraph), W3 (bgz17), W4 (mul.rs follow-
up, P0), W5 (thinking-engine VNNI).
4. CLAUDE.md § Knowledge Base (+1 row)
Inventory entry for the new knowledge doc per Mandatory
Board-Hygiene Rule.
Notable architectural decisions captured (so workers don't re-derive):
- The "alien magic" shape is struct methods + closure-batch primitives,
NOT free functions or consumer-side traits. The polyfill is the
single channel; consumers compose via closures.
- Direction B for codex P2 i8::MIN: scalar is buggy
(`unsigned_abs() as i8` wraps i8::MIN → -128), AVX-512 is correct
(`_mm512_abs_epi8` saturates to 127 by ISA). Per spec line 233 of
pr-sprint-13-simd-i4.md: |signed_mantissa| ≤ 1 → ValleyOfDespair
means "weak rule signal", not "sign-extreme". Verdict from
PP-16 preflight-drift-auditor 2026-05-16.
- Narrow scope for mul.rs follow-up + 4 separate sweep PRs, NOT one
mega-sweep. Each violator has a distinct missing-primitive
blocker; bundling would conflate 4 unrelated correctness reviews.
Per PP-14 convergence-architect SYNERGY 3 verdict.
- sigker positioning: Index-regime third codec lane (alongside
bgz17 palette-distance and deepnsm NSM tiling). Bypasses Jirak
noise floor via Hambly-Lyons 2010 uniqueness. Activates as a
first-class W1.5 consumer when jc Pillar 11 trips. Zero raw
intrinsics today — cleanest exemplar of "domain crate composes
via closures" pattern.
Pre-spawn verdict from simd-savant: the original "narrow scope"
plan was insufficient given the audit; the ndarray-first wave is
now mandatory (not just preferred). Workers W1a-#1 through W1a-#5
can spawn in parallel against adaworldapi/ndarray master once this
capture commit lands.
https://claude.ai/code/session_01UwJuKqP828qyX1VkLgGJFS
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1141770710
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| Five small primitives, each on its own branch, each auditable by `simd-savant` before merge: | ||
|
|
||
| 1. **`TD-NDARRAY-SIMD-UNPACK-I4-16D`** — `I8x16::from_i4_packed_u64` + `I8x16::lane_i8::<N>` + `batch_packed_i4_16<E, F>` closure-batch entry. AVX-512 path via `_mm512_cvtepi8_epi16` + nibble shuffle; NEON via `vshl_n_s8` / `vqshl_n_s8`; scalar via fused-loop fallback. Bounds-aware tail. | ||
| 2. **`TD-NDARRAY-SIMD-SATURATING-ABS-I8`** — `I8x16::saturating_abs(self) -> Self`. AVX-512 `_mm512_abs_epi8` (saturates `i8::MIN → 127` by ISA); NEON `vqabsq_s8`; scalar `i8::saturating_abs`. Closes codex P2 i8::MIN divergence on PR #398 by giving consumers a single source-of-truth for "the abs that matches hardware semantics." |
There was a problem hiding this comment.
When a W1a worker follows this canonical plan to implement I8x16::saturating_abs, using _mm512_abs_epi8 does not saturate i8::MIN to 127: VPABSB leaves the 0x80 lane value as 0x80, which is still -128 when represented as i8. That means the documented fix for the PR #398 i8::MIN divergence would preserve the bad AVX-512 behavior instead of matching i8::saturating_abs/NEON vqabsq_s8; the primitive needs an extra clamp/remap of 0x80 to 0x7f or the docs should not call this instruction saturating.
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Architectural capture commit per autoattended-multiagent-pattern §3 ("P0 fixes land BEFORE the next sprint PR opens"). The simd-savant PRE-MERGE audit of
origin/mainsurfaced 158 raw-intrinsic violations across 5 consumer crates plus 3 missing primitives inndarray::simdthat block clean remediation. Before any worker spawns to fix PR #398's codex P1/P2 findings, the architectural shape needs writing so every wave-1 worker briefs against the same canonical reference.What lands here (no code, only docs + board)
1.
.claude/knowledge/ndarray-vertical-simd-alien-magic.md(NEW, 144 LOC)The canonical reference. The Click: ndarray ships struct methods on typed wrappers (
I8x16,U8x32,F32x16,U64x8, …) plus closure-parameterized batch primitives; consumers compose with domain enums via closures. Per-workload surface table covers palette L1-L4, spatial splat, blasgraph-over-palette, i4 packed qualia, hamming, signature kernels.Why this shape: struct methods over free fns (surface fragments) and over consumer-side traits (single-channel property lost). The "alien magic" is that ndarray is designed AS-IF it had clairvoyant knowledge of our exact workloads — the polyfill is the cellular chemistry, consumer crates assemble organs from struct methods.
Wave plan:
I8x16::from_i4_packed_u64+batch_packed_i4_16<E,F>;I8x16::saturating_abs(codex P2 Direction B);U16x8::gather_u16; cross-archprefetch_read_t0;U64x8::popcntjc Pillar 11activation): signature-PDE-sweep, randomized-projection, lyndon-pack — forcrates/sigker2.
.claude/board/EPIPHANIES.md— PREPENDE-SIMD-SWEEP-1Full AP-SIMD-N breakdown (117 + 8 + 13 + 7 + 19 + 0 + 2 + 13 = 158). Doctrinal counterpart to
E-META-10/I-LEGACY-API-FEATURE-GATED(same retroactive-sweep shape). Strategic angle: sigker as Index-regime third lane bypassingI-NOISE-FLOOR-JIRAK.3.
.claude/board/TECH_DEBT.md— PREPEND 13 entriesTD-NDARRAY-SIMD-{UNPACK-I4-16D, SATURATING-ABS-I8, GATHER, PREFETCH, POPCOUNT-U64}— each with required API surface for the corresponding ndarray PRTD-NDARRAY-SIMD-{SIGNATURE-PDE-SWEEP, RANDOMIZED-PROJECTION, LYNDON-PACK}TD-SIMD-SWEEP-{W1 holograph, W2 blasgraph, W3 bgz17, W4 mul.rs P0, W5 thinking-engine}4.
CLAUDE.md§ Knowledge Base (+1 row)Inventory entry for the new knowledge doc per Mandatory Board-Hygiene Rule.
Architectural decisions captured (so workers don't re-derive)
unsigned_abs() as i8wrapsi8::MIN → -128), AVX-512 is correct (_mm512_abs_epi8saturates to 127 by ISA). Per spec line 233 ofpr-sprint-13-simd-i4.md. Verdict from PP-16 preflight-drift-auditor.jc Pillar 11trips.Pre-spawn verdict
simd-savant says the original "narrow scope" plan was insufficient given the 158-violation audit; the ndarray-first wave is now mandatory, not preferred. Workers W1a-#1 through W1a-#5 can spawn in parallel against
adaworldapi/ndarray masteronce this capture commit lands.https://claude.ai/code/session_01UwJuKqP828qyX1VkLgGJFS
Generated by Claude Code