[kixelated]

Finishes quest/m4/bench-moq-pin.md, planned in moq-dev/rust2#631 from moq-dev/rust2#615's report.

Problem

The moq corpus pinned f36c848, from before poll-first (moq-dev/moq#7607). The benchmark was measuring a moq its developers no longer edit: the hub edit rebuilt 8 crates instead of main's 10, and the async body edit hit a pre-poll-first async fn.

Approach

• Pin moq-dev/moq@142b03b, moq main on 2026-11-07. The leaf, hub, sync, and async targets still match verbatim. Session::closed is now kio::wait(|waiter| self.poll_closed(waiter)).await, and the comment says so.
• Drop moq-pure-crypto and mark moq as no-c. Since the pin, moq main has moved moq-relay's crypto onto rind, its own pure-Rust fork of graviola, through rind-rustls, rind-rcgen, and rind-jsonwebtoken. At 142b03b, moq-relay's build graph has no aws-lc-sys, ring, cc, cmake, or links key. So moq itself is now the pure-crypto build, and the comparison the spike corpus existed for (aws-lc-rs against pure Rust) no longer has two sides at main. The no-c guard moves to moq, and the workflow's moq pure-crypto job goes with the corpus.
• Rebuild the test list from the workspace crates moq-relay builds at the pin. moq-clock and moq-flate are gone. kio-macros, moq-congestion, and moq-crypto are new. kio no longer needs its version. moq-pattern joins moq-auth as unnameable: moq-auth's dev-dependency on its published self brings a published moq-pattern at the same version, so cargo pkgid -p moq-pattern@0.1.2 is ambiguous. All 12 remaining specs resolve to workspace crates (cargo tree -p ... at the pin).
• Keep the cranelift variant building moq. rind's naked_asm! takes sym operands, which cg_clif cannot compile yet (rust-lang/rustc_codegen_cranelift#1620, open). The first bench run at the new pin failed there. A variant gains a config key of cargo --config values, and cranelift picks its backend by profile, with LLVM for rind alone. A RUSTFLAGS backend can't do this: cargo passes RUSTFLAGS after the profile's per-package flag, so it would win. Each cargo command takes the values last: cargo drops a --config given before an external subcommand, so a leading one would have left nextest building with LLVM. The Cranelift retrial quest reads these moq rows, so dropping moq from the variant would have lost its signal.
• README: the corpus notes, the edit-async note, the Cranelift note, and the ra-jit notes, taken from the run at the new pin. ra-jit serves every edit-leaf, edit-hub, and edit-sync run. It hands every edit-async run to cargo on usage-argv's parsed_values_given, at 50.6 s and 196 CPU-s against cargo's 10.5 s and 14.5 CPU-s. The aws-lc-sys fallback note goes with aws-lc. The cargo ra-jit check note is dated to f36c848, since nothing reruns it.

Impact

• No public API. just bench --corpus moq-pure-crypto is now an unknown corpus.
• Bench numbers on either side of this PR do not compare, as corpus.toml asks. The bump lands on its own.

Alternatives

• The quest's recommendation: merge moq main into a new spike/pure-crypto branch in moq-dev/moq and pin that. I tried it. Merging main into the spike conflicts in 60 files across 1478 commits, because main replaced the same crypto code the spike patched (graviola became rind, built-in ACME was removed in moq-dev/moq#7803). Even resolved, it would reproduce what main already ships. I pushed nothing to moq-dev/moq.
• Keep an aws-lc-rs variant of moq at main for a C-vs-Rust comparison. That would mean a new spike reverting moq's rind move. moq has already made that call, so the comparison has no decision left to inform.

Choices made without asking

• Took the alternative above over the quest's recommended spike branch. The quest's premise, that the bump leaves moq-pure-crypto comparing a different tree, no longer holds, since moq is that tree.
• Kept moq on the cranelift variant with rind on LLVM, rather than dropping moq from it until rust-lang/rustc_codegen_cranelift#1620 lands. The moq Cranelift numbers now include one crate built with LLVM, which is how a Cranelift user would build moq today.
• Dropped moq-pattern's tests from test-* rather than naming it by a path+file:// spec, which would depend on the checkout path.

Validation

• nix develop --command just check: passed.
• nix develop --command just ci: only the two known path-dependent TypeId snapshots fail (lang/dyns ids.ts, lang/panics/hook lib.ts). They fail the same way on bare origin/main, and moq-dev/rust2#633 fixes them.
• The bench at 010a935 (run): moq stable, moq ra-jit, and pure passed, so the no-c guard and the body edits hold at the new pin. moq nightly failed in the cranelift warm-up on rind, which led to the fix above.
• Locally, the new config path on regex: just bench --corpus regex --variant nightly --variant cranelift --runs 1 passed, and cargo -v with the same values shows -Z codegen-backend=cranelift on every regex crate, with no warning for the unmatched rind. On moq's rind alone, rind gets llvm and its dependencies and build scripts get cranelift.
• The bench at 5015a6e: PENDING

Follow-ups

None.

Models

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

(Written by Claude Opus 5.5 with high effort)

🤖 Generated with Claude Code

Review

Rewritten by quest review after each round.

• Reviewer: Codex, gpt-6.1-sol, medium effort
• Session: 01a1b524-ea63-75c2-8a20-9cad88c366a6, round 2
• Head: 5015a6e
• Base: origin/main at efa6300
• Score: 10/10, The head fixes configuration forwarding in every affected command path, and the full branch diff reveals no remaining actionable defects.
• Findings: none
• The last round's findings, at 75e3896:
  • Fixed: high, bench/build/run.sh:247: [P1] Forward variant configuration to nextest itself; bench/build/run.sh:327, 416, and 464 now append variant configuration after the command, including nextest and test-edit preparation; the build-platforms reproduction confirms nextest receives configuration in this position.

[pullfrog]

  ^{｜ Review this ➔}