Feature: wp38-s3a-cpu-flag #657

Activity

1. amritk

   added

   feature-run

   on Oct 16, 2026
2. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log, check-in #1 (09:43Z):

   • s3a-measure: PR perf(bench): measure S3's bar kernels without a nish:simd surface (WP38 S3a) #658 @ 526652b, CI green (21 checks; the first no attribution run failed on the cloud footer and passed on re-run), clean against main @ a22a056 → review round 1 started (2 shards)
   • s7-other-holder-golden: PR test(checker): pin S7's guard-on-another-array negative golden #659 opened, CI pending; subscribed
   • s2-cpu-flag: worker running (goldens regenerated, reviewing its diff), no branch yet; no-branch clock 09:57Z

   ---

   Generated by Claude Code
3. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 09:5xZ: #658 r1 @ 526652b. 11 raised (2 shards), 0 blocking confirmed → converged. 7 non-blocking findings sent once (§7 row "landed" wording, lscpu detail, the form-2-vs-3 gap, the bool-vector bitmask note, --only "", --runs rounding, the source regex). The worker was told: one push, no further round, the lead merges on green.
   Rejected (#658 r1): the remark-based vectorisation report (deviation recorded in the record; the default profile is speed); the vectorised() crash on remark paths (no reachable input); dropping ,/: from the depth kernel (the plan names four byte classes); the "3 over its copy" row being inverted (it follows the table's X-over-Y convention).

   ---

   Generated by Claude Code
4. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 09:5xZ:

   • perf(bench): measure S3's bar kernels without a nish:simd surface (WP38 S3a) #658: the worker pushed one commit applying all 7 non-blocking findings (§7 row "Landed (perf(bench): measure S3's bar kernels without a nish:simd surface (WP38 S3a) #658)" naming 46a1e4d, lscpu detail, the form-2-vs-3 gap stated, the bitmask note, --only "" → exit 2, --runs rounding documented, the regex widened). Waiting on CI at that head.
   • test(checker): pin S7's guard-on-another-array negative golden #659 r1 @ bb6aefb: 0 findings raised beyond one prose nit (the offset-index block's header comment names three rules, not the new fourth; not sent, since it is not worth a CI cycle). Converged. The gate passed: CI 21/21 · clean · diff ⊆ Owns · no sensitive paths · first in merge order. Squash-merged as 917897c. Unsubscribed; s7 worker archived.
   • s7-other-holder-golden: merged. s2-cpu-flag is next in merge order once it converges; it merges main and regenerates tests/self/goldens first.

   ---

   Generated by Claude Code
5. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 10:0xZ:

   • perf(bench): measure S3's bar kernels without a nish:simd surface (WP38 S3a) #658: CI 19/19 green @ cfdc8ef. Gate passed (converged r1 · nothing skipped/disputed · clean · diff ⊆ Owns · no sensitive paths · no merge-order dependency) → squash-merged as b8c3b72 with the changelog body and the Measured: trailer from the record. Unsubscribed; s3a worker archived. s3a-measure: merged.
   • s2-cpu-flag: PR feat(cli): --cpu x86-64-v3 links module and runtime at one CPU level (WP38 S2) #660 has been open since 09:50Z @ c52853d (branch pushed before the 09:57Z clock). CI is green (28 checks; the first no attribution run failed on the cloud footer and was superseded by a success). No textual conflict with main @ b8c3b72. Subscribed → review round 1 started (2 shards: src/build.sh/ci.yml; tests/docs).

   ---

   Generated by Claude Code
6. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 10:1xZ: #660 r1 @ c52853d. ~12 raised (2 shards), 7 verified, 0 blocking → converged. 5 non-blocking findings were sent once, together with the "you are next" instruction: merge main (S7 + S3a landed), regenerate goldens, one push. The findings: the CT check counts, the CT-16 README index row, the stale "not built" lines, LANGUAGE.md's host sentence, the README usage line.
   Owns amended by the lead for S2 (doc pointers to the same fact; no other stage is open): the CT-16 row of docs/security/README.md, the S2 "not built" status lines of docs/wp38-simd.md, and one README.md usage line.
   Rejected (#660 r1): CT-16's finding column vs its status column (the file's convention); the guide's "Haswell … and later" claim (conditioned on AVX2). Not sent, observations only: --cpu without --link is a no-op; the "-march on 7 units" and cache checks have no CI grep (they are not /proc/cpuinfo-gated).

   ---

   Generated by Claude Code
7. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 10:4xZ: #660's worker pushed 63991c1. It merged main @ 464f41e (#642) and applied the 5 non-blocking findings plus the authorised pointer edits; CI completed, clean. Main then moved to b322a4b (#652, std/ and goldens), so per CLAUDE.md the worker was told to merge main again and regenerate before the gate. Recorded deviation: the capability report writes "cpu" only with the flag; a baseline report stays byte-identical rather than carrying "cpu": "baseline". The PR body says this avoids moving the caps goldens and the IR cookbook, which are outside Owns. That bears on acceptance criterion 4 and is left for the owner.

   ---

   Generated by Claude Code
8. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Log 10:58Z:

   • feat(cli): --cpu x86-64-v3 links module and runtime at one CPU level (WP38 S2) #660 @ f2a37dc: the worker merged main @ b322a4b (fix(tls): wipe every key a TLS connection holds when it ends #652). Regenerating the goldens left them unchanged. CI 26/26 green. The gate passed: converged r1 · nothing skipped or disputed · the branch contains main · diff ⊆ amended Owns · ci.yml has only the authorised grep step · s7 is merged. → squash-merged as 3756681. Unsubscribed; s2 worker archived. s2-cpu-flag: merged.
   • All 3 slices merged. Phase 7 has started: the static pass (in-process) and the acceptance session session_01HFtMokKdVhcZbZJDH3zvz3 (sonnet; CLI exercise on main @ 3756681; verdict comment behind <!-- feature:acceptance -->).

   ---

   Generated by Claude Code
9. amritk commented on Oct 16, 2026

   

   amritk

   on Oct 16, 2026 – with Claude

   OwnerAuthor

   More actions

   Verdict: FAIL. Seven of the eight criteria pass. Criterion 1 fails on its "under Node" half. Native behaves as specified, but the same program under Node prints NaN and exits 0.

   Tested at 375668139fbd (detached checkout of main) after bash scripts/fetch-seed.sh --force (exit 0) and npm run build (exit 0). The machine has AVX2: grep -c avx2 /proc/cpuinfo printed 4. It is an Intel Xeon @ 2.80GHz with clang 18.1.3.

   #	Criterion	Result	Evidence
   1	arr_bounds_offset_other panics index out of range: 2 >= 2, exit 1, native and Node; mixed keeps one nish_panic_index branch, on ys	FAIL (native half and IR pass)	Native: nish tests/cases/arr_bounds_offset_other.ts --link then run prints index out of range: 2 >= 2, exit 1. nish run gives the same, exit 1. IR: tests/cases/arr_bounds_offset_other.ll has one call @nish_panic_index(i64 %26, i64 %5). %5 is the length loaded from %ys, and xs has no check. Node: node --experimental-strip-types --import ./runtime/nish.mjs -e 'const m = await import("./tests/cases/arr_bounds_offset_other.ts"); process.exit(m.main())' prints NaN, exit 0. runtime/nish.mjs:29-30 documents that a[i] is unchecked under Node, and tests/differential/goldens/unfrozen.txt:28 lists the case as UNFROZEN, so no run compares Node with native.
   2	node bench/simd-s3a.mjs --self --runs 15	pass	Exit 0. It built all four forms of mat-vec and depth, with one checksum per kernel (901781, 680640 and 176206 for depth s0 and log). It printed min and median for each form and the noise-band row. It printed portability: no target intrinsic in any vector form's source or IR (the x86-64 and AArch64 probes are caught). bench/simd-s3a.mjs:139-153 runs the _mm_sad_epu8 and vqtbl1q_u8 probes and fails if either is missed. --only "" prints --only needs at least one kernel (kernels: mat-vec, depth) and exits 2. I did not plant a refusable vector file.
   3	docs/wp38-simd.md §2.3 S3a record; §7 row landed	pass, with a caveat	§2.3 (lines 651-785) has the commit, the machine, clang, the three tables, the ratios against the 3 over its copy noise band, a section on whether the scalar forms vectorised, the AArch64 line ("no AArch64 figure") and the reading. The §7 S3a row (line 1325) says "Landed (#658)". Caveat: the cited commit 46a1e4d is not an object in this checkout. It was a pre-squash PR commit, so the record's anchor can't be resolved from main.
   4	--cpu x86-64-v3 --link on examples/hello.ts	pass	Baseline and v3 builds both exit 0 and print hello from Nish, and cmp of the outputs is identical. --emit-capabilities with the flag has "cpu": "x86-64-v3" right after "entry". The baseline report has no cpu key (grep -c '"cpu"' gives 0). Its top-level keys are version, entry, deterministic, capabilities and packages.
   5	Five refusals exit 2 with one stderr line naming x86-64-v3	pass	Each exit was 2 with one stderr line. --cpu native, --cpu x86-64-v4 and --cpu armv8.2-a each print compile: unknown cpu level `<x>`; the one accepted is x86-64-v3. --cpu x86-64-v3 --profile wasi prints compile: --cpu x86-64-v3 is an x86-64 level, so it cannot be used with --profile wasi. --cpu x86-64-v3 --target aarch64-unknown-linux-gnu prints compile: --cpu x86-64-v3 is an x86-64 level, and the target is aarch64-unknown-linux-gnu.
   6	With and without --cpu against one cache gives two runtime entries	pass	With XDG_CACHE_HOME set to a fresh directory, nish run with and without --cpu left 2 dirs in nish/runtime. Repeating them left it at 2. The same with nish --link on a fresh cache, run three times (baseline, v3, baseline), also left 2. Mixing run and --link in one cache gave 4, so the two routes keep separate entries.
   7	Every ct_asm check passes at baseline and x86-64-v3	pass	node tests/run.js ct_asm gave 222 passed, 0 failed, exit 0, with no skip lines and 0 FAIL lines. 58 lines name x86-64-v3, including the fixtures (ct_asm_aes, ct_asm_chacha20poly1305) and the runtime-aes.c kernels (exports, nish_aes_schedule, nish_aes_block).
   8	Docs describe the flag; links pass	pass	docs/LANGUAGE.md has "The CPU level" (line 5014, and 7539 onward). docs/guide/using-nish/cli.md:78 has the flag. CT-16 in docs/security/ct-verification.md:89 is marked "Narrowed (WP38 S2)". node docs/check-links.mjs printed ok: 1346 links in 87 Markdown files, exit 0.

   Other commands.

   • npm run check: exit 0.
   • npm run lint: exit 0 (Checked 444 files).
   • node tests/run.js cpu-level: 31 passed, 0 failed, exit 0.
   • node tests/run.js ct_asm: 222 passed, 0 failed, exit 0.
   • node tests/run.js arr_bounds_offset: 51 passed, 0 failed, exit 0, no skip lines.

   Observation, not a criterion. In my --self run the depth s0 noise band was 1.11x min / 1.08x median, against "within 5%" in the §2.3 record. The C vector forms still beat Nish checked by 1.61x on mat-vec, 2.16x on depth s0 and 5.84x on depth log (min), so the bar still holds on this machine.

   Not exercised.

   • The Node half of criterion 1 as specified. Node cannot produce the panic, per the documented behaviour above.
   • The AArch64 line. There is no AArch64 machine, and the record itself says there is no AArch64 figure.
   • Planting a file that holds llvm.x86. or llvm.aarch64. to see the refusal. I relied on the harness's own two probes.
   • The full npm test, as instructed. Nothing was fixed, and no PR, branch or issue was opened.

   ---

   Generated by Claude Code
10. amritk commented on Oct 16, 2026

    

    amritk

    on Oct 16, 2026 – with Claude

    OwnerAuthor

    More actions

    Close-out: 3 merged, 0 escalated slices; acceptance 6/8, 2 criteria escalated to #662

    stage	PR	state
    s7-other-holder-golden	#659	merged as 917897c (r1, converged)
    s3a-measure	#658	merged as b8c3b72 (r1, converged; 7 non-blocking applied)
    s2-cpu-flag	#660	merged as 3756681 (r1, converged; 5 non-blocking applied; main merged twice before the gate)
    acceptance	—	escalated: static pass and session session_01HFtMokKdVhcZbZJDH3zvz3

    Review: 3 rounds in all. 0 blocking findings were confirmed, 12 non-blocking were applied and 6 were refuted.
    CI on main @ 3756681: the CI, Interop and Release PR workflows all succeeded.

    Acceptance: criteria 2, 3, 5, 6, 7 and 8 are met. The criteria were derived from the plan, not given separately.

    • 1 (escalated). arr_bounds_offset_other panics natively but prints NaN under Node. Indexing is unchecked under Node by design, and cases added after R6 are not compared against Node, so the plan's expectation was wrong rather than the code.
    • 4 (escalated). A baseline capability report has no cpu key; the plan asked for "cpu": "baseline". feat(cli): --cpu x86-64-v3 links module and runtime at one CPU level (WP38 S2) #660 departed from the plan deliberately and documented it.
      Both are in WP38 S7/S2 acceptance: two criteria the merged code does not meet as written #662 for the owner. This issue stays open as the handoff.

    Proposed CLAUDE.md line. #658 and #660 each anchored a record at a branch SHA that the squash merge made unreachable:

    A record or §7 row that anchors at a commit names one that survives the squash merge: the PR number, or the squash SHA added in the merge commit's own record update. A branch SHA is unreachable from main once the PR is squashed.

    Report: https://claude.ai/artifact/7WLnAi9jBv4khrMau59r4k

    ---

    Generated by Claude Code