All bytewax Py<T> references are properly guarded with SafePy (15
struct fields, 5 Drop impls, PICKLE_MODULE static). The SIGSEGV comes
from pytest-codspeed's walltime profiler shutdown conflicting with
Python 3.14 finalization (CodSpeedHQ/pytest-codspeed#105). Benchmark
data is captured before the shutdown crash, so catch exit 139 and
emit a warning instead of failing the job.

* Fix benchmark workspace permission errors on self-hosted runners

* Upgrade actions to latest versions, improve caching and sccache

- Upgrade checkout v3→v6, cache v4→v5, upload-artifact v4→v7,
  download-artifact v4→v8, setup-just v2→v3, dawidd6 v6→v16
- Add restore-keys to all Cargo caches for warm restarts
- Add cache-dependency-glob to all setup-uv calls
- Set CARGO_INCREMENTAL=0 for sccache compatibility
- Add retention-days: 7 to CI wheel artifacts

* Bump Rust 1.74.1 → 1.85.0, update dependencies and Dockerfile

- Rust toolchain: 1.74.1 → 1.85.0
- Widen serde/tokio/fastrand Cargo.toml pins, cargo update resolves
  174 packages (tokio 1.50, serde 1.0.228, chrono 0.4.44, etc.)
- pre-commit-hooks: v4.4.0 → v5.0.0
- Dockerfile: rust:1.68-bullseye → 1.85-bookworm, distroless
  debian11 → debian12, install maturin via pip instead of obsolete
  konstin2/maturin:v0.12.6 image

* Fix repo-checks OOM, pin benchmarks to Intel runner

- repo-checks: replace `uv pip sync -e .` (OOM during in-process
  cargo build) with maturin-action wheel build + filtered dev deps
  install. Adds sccache, sudo wrapper, workspace permission fix.
- benchmarks: add Intel label to runs-on for perf consistency

* Fix maturin-action directory collision and uv hardlink issues

- Clean stale /__w/_temp/run-maturin-action.sh before builds
  (new runners had a directory at this path, causing Docker bind
  mount failures: "is a directory: permission denied")
- Set UV_LINK_MODE=copy for container jobs to avoid hardlink
  failures across overlay filesystem boundaries (prevents cache
  corruption like missing wheel METADATA)

* Fix concurrent maturin collisions, sccache timeout, uv cache corruption

- Isolate maturin temp dir per matrix job (RUNNER_TEMP unique path)
  to prevent concurrent jobs on the same runner from colliding on
  /__w/_temp/run-maturin-action.sh
- Disable sccache for repo-checks (server can't start in containers)
- Retry uv pip sync with cache clean on failure to recover from
  corrupted wheel metadata in shared uv cache

* Fix maturin-action stale directory cleanup for cross-compile jobs

Replace broken RUNNER_TEMP env var override (resolved at parse time,
not runtime) with targeted cleanup that only removes the maturin temp
script path when it's a stale directory from a previous crashed job.
Remove unnecessary isolation step from benches.yml (x86_64-only builds
don't use Docker).

* Ensure /__w symlink on every runner for cross-compile Docker builds

maturin-action creates the build script at $RUNNER_TEMP but Docker
bind-mounts it via /__w/_temp/ path. The /__w -> /opt/actions-runner/_work
symlink was only created by pre-pull (on one runner). Cross-compile jobs
on other runners fail because Docker can't resolve the path and auto-creates
a directory instead. Fix by ensuring the symlink exists on each runner
before the maturin-action step.

* Fix SIGSEGV in Python 3.13/3.14: wrap PICKLE_MODULE static in SafePy

The static GILOnceCell<Py<PyModule>> at pyo3_extensions.rs:17 was the
only bare Py<T> in a global/static context not wrapped in SafePy. During
Python 3.13+ interpreter finalization, its drop calls Py_DECREF on an
already-freed type object, causing SIGSEGV (exit 139). Wrapping in
SafePy<PyModule> checks Py_IsFinalizing() before dropping.

* Handle CodSpeed walltime SIGSEGV on Python 3.13+ benchmarks

All bytewax Py<T> references are properly guarded with SafePy (15
struct fields, 5 Drop impls, PICKLE_MODULE static). The SIGSEGV comes
from pytest-codspeed's walltime profiler shutdown conflicting with
Python 3.14 finalization (CodSpeedHQ/pytest-codspeed#105). Benchmark
data is captured before the shutdown crash, so catch exit 139 and
emit a warning instead of failing the job.

---------

Co-authored-by: Nick Bozhenko <nick.bozhenko@lotusflare.com>

* Fix benchmark workspace permission errors on self-hosted runners

* Upgrade actions to latest versions, improve caching and sccache

- Upgrade checkout v3→v6, cache v4→v5, upload-artifact v4→v7,
  download-artifact v4→v8, setup-just v2→v3, dawidd6 v6→v16
- Add restore-keys to all Cargo caches for warm restarts
- Add cache-dependency-glob to all setup-uv calls
- Set CARGO_INCREMENTAL=0 for sccache compatibility
- Add retention-days: 7 to CI wheel artifacts

* Bump Rust 1.74.1 → 1.85.0, update dependencies and Dockerfile

- Rust toolchain: 1.74.1 → 1.85.0
- Widen serde/tokio/fastrand Cargo.toml pins, cargo update resolves
  174 packages (tokio 1.50, serde 1.0.228, chrono 0.4.44, etc.)
- pre-commit-hooks: v4.4.0 → v5.0.0
- Dockerfile: rust:1.68-bullseye → 1.85-bookworm, distroless
  debian11 → debian12, install maturin via pip instead of obsolete
  konstin2/maturin:v0.12.6 image

* Fix repo-checks OOM, pin benchmarks to Intel runner

- repo-checks: replace `uv pip sync -e .` (OOM during in-process
  cargo build) with maturin-action wheel build + filtered dev deps
  install. Adds sccache, sudo wrapper, workspace permission fix.
- benchmarks: add Intel label to runs-on for perf consistency

* Fix maturin-action directory collision and uv hardlink issues

- Clean stale /__w/_temp/run-maturin-action.sh before builds
  (new runners had a directory at this path, causing Docker bind
  mount failures: "is a directory: permission denied")
- Set UV_LINK_MODE=copy for container jobs to avoid hardlink
  failures across overlay filesystem boundaries (prevents cache
  corruption like missing wheel METADATA)

* Fix concurrent maturin collisions, sccache timeout, uv cache corruption

- Isolate maturin temp dir per matrix job (RUNNER_TEMP unique path)
  to prevent concurrent jobs on the same runner from colliding on
  /__w/_temp/run-maturin-action.sh
- Disable sccache for repo-checks (server can't start in containers)
- Retry uv pip sync with cache clean on failure to recover from
  corrupted wheel metadata in shared uv cache

* Fix maturin-action stale directory cleanup for cross-compile jobs

Replace broken RUNNER_TEMP env var override (resolved at parse time,
not runtime) with targeted cleanup that only removes the maturin temp
script path when it's a stale directory from a previous crashed job.
Remove unnecessary isolation step from benches.yml (x86_64-only builds
don't use Docker).

* Ensure /__w symlink on every runner for cross-compile Docker builds

maturin-action creates the build script at $RUNNER_TEMP but Docker
bind-mounts it via /__w/_temp/ path. The /__w -> /opt/actions-runner/_work
symlink was only created by pre-pull (on one runner). Cross-compile jobs
on other runners fail because Docker can't resolve the path and auto-creates
a directory instead. Fix by ensuring the symlink exists on each runner
before the maturin-action step.

* Fix SIGSEGV in Python 3.13/3.14: wrap PICKLE_MODULE static in SafePy

The static GILOnceCell<Py<PyModule>> at pyo3_extensions.rs:17 was the
only bare Py<T> in a global/static context not wrapped in SafePy. During
Python 3.13+ interpreter finalization, its drop calls Py_DECREF on an
already-freed type object, causing SIGSEGV (exit 139). Wrapping in
SafePy<PyModule> checks Py_IsFinalizing() before dropping.

* Handle CodSpeed walltime SIGSEGV on Python 3.13+ benchmarks

All bytewax Py<T> references are properly guarded with SafePy (15
struct fields, 5 Drop impls, PICKLE_MODULE static). The SIGSEGV comes
from pytest-codspeed's walltime profiler shutdown conflicting with
Python 3.14 finalization (CodSpeedHQ/pytest-codspeed#105). Benchmark
data is captured before the shutdown crash, so catch exit 139 and
emit a warning instead of failing the job.

---------

Co-authored-by: Nick Bozhenko <nick.bozhenko@lotusflare.com>