[kixelated]

Works on quest/m1/bench-apple-m4.md: adds kixtop (Apple M4, macOS) to BENCHMARKS.md, and rewrites the quest to what remains, a quiet desktop re-collect and kixtop's build cost, both waiting on the maintainer's hosts going quiet.

Changes

• benchmarks/data/kixtop.json: the second of two native kixtop collections, at 93f91f48, which measured 0.68 other busy CPUs during the run (not noisy). The first, at cc2f73f6, ran under 4.36 other busy CPUs and most rows came out 30-40% slower, so it is not used. fix(bench-report): hold each host with quest gate --host, and measure macOS noise #399's macOS path worked unchanged; the ssh host key was already accepted.
• benchmarks/gaps.toml: two kixtop entries, both reasons:
  • 19 short-message and packet-size rows (AES-GCM 32 B and 1350 B, ChaCha20-Poly1305 32 B, all three header protections, SHA-256/512, HMAC, HKDF), 1.11x to 12.6x: DIT. rind toggles it around every call that handles a secret, about 35-55 ns a call on the M4, and no other library sets it. A probe build on kixtop at the same commit with DIT denied (--cfg rind_cpu_disable, RIND_CPU_DISABLE_dit) put every one within 10% of the fastest library except AES header protection (1.2x, 0.6 ns) and HMAC-SHA-256 64 B (1.15x RustCrypto). Per-call DIT is the maintainer's decision from #387's findings, and #431's rind::cpu::Dit::hold() pays it once per batch.
  • xchacha20poly1305/32B 1.26x: noise. The probe's rerun and the first run both measured 1.04x.
• admin/bench-report: a kixtop host note (in daily use, has DIT), and "in the Nix dev shell on Linux" becomes "in the Nix dev shell", since macOS uses it too since #399.
• README.md: the gaps example adds AES-GCM and short messages on Apple silicon, so "matches or beats both" stays true with kixtop in the headline.
• quest/m1/bench-apple-m4.md: rewritten to the remaining work, with a Waiting item for the maintainer. quest mv refuses on a claimed quest, so it keeps its path. quest/m3/bench-report-refresh.md now points here instead of at the quest.

Verified table

Rechecked against kixtop's features (aes, dit, neon, pmull, sha2, sha3); no change needed. With sha3, rind runs the SHA-512 instructions, so the !sha3 SHA-512 and Ed25519 entries correctly give no ✓. X25519, P-256 ECDH and ECDSA, and ChaCha20 header protection run the _alt routines and the same Rust their aarch64 entries cover, so they keep theirs.

Why desktop isn't here

collect refused desktop three times on 2026-11-07 (16.8, 4.7, and 5.3 of 24 CPUs busy, limit 2.4), all from other sessions' ungated builds. One attempt's host hold waited an hour: the global quest on desktop predates moq-dev/quest#526 and has no host-hold logic, so its gated commands were admitted past the hold. Both kixtop runs' build costs were refused the same way (11.6 and 6.6 of 10 busy), so its build-cost table reads "Not measured".

Choices made without asking

• Reported kixtop's quiet second run rather than averaging in the noisy first.
• Recorded the DIT rows as a reason, not a quest, since keeping DIT per call is already settled.
• Landed kixtop now and moved desktop to the quest instead of holding the PR open for a quiet window that may not come soon.

Impact

No public API or wire change. Report data, its renderer's host note, README wording, and quests only.

Validation

• nix develop --command just check origin/main: pass on the pushed head.
• admin/bench-report render --check: pass.
• quest check --base origin/main: 70 documents ok.

Follow-ups

• Time rind under rind::cpu::Dit::hold() on DIT cores in the report or nightly bench. feat(cpu): hold DIT across a batch of calls, putting M1 AES-GCM ahead of aws-lc-rs #431 skipped a held case because no report host had DIT; kixtop does, and MoQ will hold the guard, so the report shows a cost MoQ won't pay. Recommend: plan it.
• aarch64 HMAC-SHA-256 at 64 B, 1.15x RustCrypto with DIT denied. Recommend: drop; it is about 15 ns, and a handshake runs only a few HMACs against 130 µs or more.
• AES header protection on aarch64, 1.2x aws-lc-rs with DIT denied. Recommend: drop; 0.6 ns a packet.

Models

Size: S
Implementer: Claude Code, claude-opus-5-5, high effort

Review

Rewritten by quest review after each round.

• Reviewer: Codex, gpt-6.1-sol, medium effort
• Session: 01a1b43d-ccb7-7090-b831-b32b0ebf82f3, round 1
• Head: b44c0c6
• Base: origin/main at e7570f6
• Score: 10/10, The specified diff adds no cryptographic paths, the report and charts pass freshness validation, all eight rendering tests pass, and the missing M4 build costs are explicitly disclosed.
• Findings: none

[pullfrog]

Your OpenCode Go account is out of credit. Top it up, then re-trigger Pullfrog.

OpenCode Go's response

provider error: Insufficient account funds

Top up OpenCode Go →

  ^{｜ Rerun failed job ➔ ｜ View workflow run ｜ via Pullfrog ｜ Using glm-5.3}

[coderabbitai]

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 5 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

Learn how review limits work.

Review configuration:

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 64155bf6-ec1b-4020-9455-b35e4a12c796

Commits

Reviewing files that changed from the base of the PR and between e7570f6 and 58faaf5.

Files ignored due to path filters: 6

• benchmarks/headline-dark.svg is excluded by !**/*.svg
• benchmarks/headline.svg is excluded by !**/*.svg
• benchmarks/kixtop-dark.svg is excluded by !**/*.svg
• benchmarks/kixtop.svg is excluded by !**/*.svg
• benchmarks/overview-dark.svg is excluded by !**/*.svg
• benchmarks/overview.svg is excluded by !**/*.svg

Files selected for processing: 9

• BENCHMARKS.md
• README.md
• admin/bench-report
• benchmarks/data/kixtop.json
• benchmarks/gaps.toml
• quest/m1/README.md
• quest/m1/bench-apple-m4.md
• quest/m1/bench-dit-held.md
• quest/m3/bench-report-refresh.md

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn

[kixelated]

Automated review by review (OpenAI)

Reviewed head b44c0c6 against base e7570f6.

No actionable findings in the reviewed report/data changes. The M4 collection identifies commit 93f91f4, records 0.68 other busy CPUs against a limit of 1, and explicitly records the failed build-cost collection rather than inventing a measurement. Its 357 result IDs are unique and each estimate lies within its recorded interval. Recomputing the >10% comparisons from those results gives exactly the 19 DIT-listed operations plus the XChaCha20 32 B row, including the reported 12.62x maximum. The M4's sha3 feature excludes the !sha3 SHA-512/Ed25519 verification entries. The remaining quiet Arrow Lake collection and M4 build costs stay open in the quest.

Direction: the README now acknowledges Apple's short-message gaps. A separate held-DIT measurement remains useful for batch consumers, but keep it distinct from this per-call baseline and from the DIT-disabled diagnostic probe. Preserving the source commit and load qualification is important when newer kernel or dispatch changes land.

Limits: GitHub-only review of the report, raw JSON, gap entries, small renderer changes, verification predicates and quest updates. I did not rerun collection, the DIT-disabled probe, renderer freshness tests, CI, or proofs, and did not visually render the SVGs. The noise and causal explanations rely on the additional runs described in the PR, not independently reproduced measurements. This COMMENT is not merge approval.