rustc: use llvmPackages_23 - #578990

…ches LLVM

rustc 1.99 in nixpkgs is built with LLVM 21 but its profiler runtime
expects the LLVM 23 raw profile format. Every coverage-instrumented
binary SIGSEGVs at exit (initializeValueProfRuntimeRecord), so
cargo llvm-cov fails. Take cargo, rustc, cargo-llvm-cov and llvm for the
tests check from a second nixpkgs input pinned at 55bfb4d. This replaces
the doCheck override.

Tracked upstream in NixOS/nixpkgs#578990.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

First use of nix-detour's fixes (kleinbem-dev/nix-detour#23). qdrant
fails to build on nixos-unstable because rustc's stdarch and nixpkgs'
LLVM disagree on an AVX-512 VNNI intrinsic (NixOS/nixpkgs#553191). It
was held on stable (1.17.1) since 10-13; now it runs the current 1.18.2
with the source patch from NixOS/nixpkgs#553347, which disables those
VNNI code paths, pinned by hash. The fix ends like a hold: once nixpkgs'
own qdrant builds again (NixOS/nixpkgs#578990 on master, plus a zstd
input that staging-next now needs), the release action removes it.

- `held-packages.json`: qdrant from `current` with the patch
(`nix-detour fix add`).
- Lock and release action move to nix-detour main (ce248312), which has
fixes; `overlay.nix` is the only file that changed for the holds.

Test (nixos-nvme, qdrant as nix-config gets it: this flake's nixpkgs,
nix-packages' overlay, `mkOverlay`): built in 91 min, 138 tests passed,
version 1.18.2; the server started, created a collection with int8
quantization, took points and answered a query. The other four holds are
unchanged.

Cost: qdrant isn't on cache.nixos.org with the patch, so nix-config's
first build-all after this builds it once (about 1.5 h here, more on a
runner); then it's in Attic. Trade-off: CPUs with AVX-512 VNNI lose that
quantization fast path until the fix ends.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
<!-- coderabbit-ai-generated:v1 -->

## Summary by CodeRabbit

* **Bug Fixes**
* Updated the Qdrant package hold to track the current package source,
with a start date of 1 November 2026. A compatibility patch disables
affected AVX-512 VNNI code paths. The hold now identifies a rustc/LLVM
compilation failure involving `vpdpbusd` and records the conditions for
ending the hold, making the package’s build status and hold criteria
clearer.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

The lock file maintenance moves nixpkgs to 4fe25e1, where cargo-llvm-cov
0.9.0 no longer builds: 14 of its integration tests fail because the
coverage-instrumented binaries they build segfault at exit. nixpkgs'
rustc is built against LLVM 21 but bundles an LLVM 23 profiler runtime
(NixOS/nixpkgs#576588, reproduced on Hydra for x86_64-linux and here on
aarch64-darwin). With the package unbuildable and not cached, the dev
shell fails to build, so direnv and the generated git hooks go stale.

The defect is in nixpkgs' rustc, not in cargo-llvm-cov, and only its
own test suite compiles with that rustc. The dev shell runs
`cargo llvm-cov` with the rust-overlay toolchain from
rust-toolchain.toml, whose profiler runtime matches its LLVM, so
skipping the package's checks yields a working tool. This is applied on
every platform because the cause is not Darwin-specific.

The fix (NixOS/nixpkgs#578990, rustc: use llvmPackages_23) is on staging;
drop the override once it reaches the pinned nixpkgs.

* chore(deps): lock file maintenance

* chore: format

* style(nu): apply nufmt 0.1.6 formatting to scripts merged from main

The lock file maintenance bumps nixpkgs from 8edc0c7 to 4fe25e1, which
moves nufmt from 0-unstable-2026-09-28 to 0.1.6-unstable-2026-10-21. The
new nufmt wraps long records and calls across lines (for example
multi-line `error make {msg: ...}` records), so treefmt-check fails on
every .nu file still in the old style.

autofix-ci had already reformatted the one script present on this
branch, but CI checks the merge with main, and main has since gained
contribution-gate and other scripts formatted by the old nufmt. That
single treefmt-check failure broke both the `security & lint preflight`
job and `nix flake check` in the `test` job; the Rust, JS, Nushell, and
Babashka tests themselves passed.

Merge main and rerun `nix fmt` so every script matches the formatter the
lock now pins.

* fix(nix): skip cargo-llvm-cov checks so the dev shell builds

The lock file maintenance moves nixpkgs to 4fe25e1, where cargo-llvm-cov
0.9.0 no longer builds: 14 of its integration tests fail because the
coverage-instrumented binaries they build segfault at exit. nixpkgs'
rustc is built against LLVM 21 but bundles an LLVM 23 profiler runtime
(NixOS/nixpkgs#576588, reproduced on Hydra for x86_64-linux and here on
aarch64-darwin). With the package unbuildable and not cached, the dev
shell fails to build, so direnv and the generated git hooks go stale.

The defect is in nixpkgs' rustc, not in cargo-llvm-cov, and only its
own test suite compiles with that rustc. The dev shell runs
`cargo llvm-cov` with the rust-overlay toolchain from
rust-toolchain.toml, whose profiler runtime matches its LLVM, so
skipping the package's checks yields a working tool. This is applied on
every platform because the cause is not Darwin-specific.

The fix (NixOS/nixpkgs#578990, rustc: use llvmPackages_23) is on staging;
drop the override once it reaches the pinned nixpkgs.

* fix(perf): read hyperfine 2.0 results by command name

The lock file maintenance moves nixpkgs to 4fe25e1, which upgrades
hyperfine from 1.21.0 to 2.0.0. The PR performance comparison then fails
before measuring anything: hyperfine 2.0 removed `--sort` without a
replacement, so `--sort command` is rejected as an unexpected argument.

Dropping the flag alone is not enough. hyperfine 2.0 also switched
`--export-json` to schema version 2: statistics moved to
`summary.time_wall_clock`, per-run times to `measurements`, and the
`--command-name` label to `name`. The parser still read the 1.x top-level
`median`/`min`/`max`/`times`, so it would have produced nil measurements.
The script also relied on `--sort command` to keep results in input
order, destructuring them by position into base/PR and pairing them with
runtime variants by index.

Match each result to its requested `--command-name` and return them in
the requested order, failing on a missing or duplicated name instead of
silently pairing the wrong measurement. Read wall-clock statistics from
the v2 summary, rejecting an unexpected schema version or unit. Remove
the unused `hyperfine-common` definition that also carried `--sort`.

The malformed-output test passed a java.nio Path to `slurp`, which
failed to open it, so it passed on the wrong exception. The test helper
now passes the string path the script itself uses.

---------

Co-authored-by: renovate[bot] <29139614+renovate[bot]@users.noreply.github.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: ryoppippi <[email protected]>

nixos-unstable rustc builds against LLVM 23 since nixpkgs#578990, so
the check that kept gitbutler below 0.23 no longer triggers.

NixOS/nixpkgs#578990