ci: run coverage via nextest to fix false tracing coverage gaps

[martintmk]

Problem

Codecov intermittently reports missing coverage on tracing::event! / log macro
lines (for example the field-value expressions in
crates/fetch_hyper/src/connection/client_connector.rs), even though those paths
are exercised by tests. Those expressions only execute on the macro's
"subscriber is interested" branch. tracing caches callsite interest
process-globally, so under the default single-process cargo test harness a
no-subscriber test can register a callsite as never-interested before a
subscriber-installing test runs. The enabled branch is then recorded as
uncovered, nondeterministically and differently per platform, producing spurious
Codecov gaps.

Change

Generate coverage with cargo llvm-cov nextest (one process per test) instead of
the default cargo test harness. Per-test process isolation removes the
cross-test contamination of tracing's callsite-interest cache, so the macro's
enabled branch is measured deterministically. This also aligns CI with the
existing just coverage-measure recipe. cargo-nextest is added to the coverage
job's tool installs.

nextest also compiles the non-test library of leaf crates that the old harness
never built, so a few pre-existing latent issues are fixed to keep the coverage
build clean under -D warnings and at 100%:

• automation / fetch_hyper enable feature(coverage_attribute) whenever
  coverage_nightly is set (they apply #[coverage(off)] to non-test items).
• Remove redundant test-only imports in fetch_hyper's hyper_handler.
• Allow the unused_features lint (the coverage-attribute gate is legitimately
  unused in feature/target combinations where a crate's coverage(off) is fully
  cfg-gated out).
• Drop cachet's uncoverable no-op logging fallbacks (let _ = (..) blocks that
  only exist in non-test, no-logs builds and can never be executed by a test).

Validated locally: llvm-cov nextest passes clean under -D warnings for both
--all-features (4752 tests) and --no-default-features (3882 tests), with no
uncovered lines.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.0%. Comparing base (76824e1) to head (193fc11).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #557   +/-   ##
=======================================
  Coverage   100.0%   100.0%           
=======================================
  Files         356      356           
  Lines       26966    26966           
=======================================
  Hits        26966    26966           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR updates the CI coverage workflow to generate deterministic coverage results by running tests under cargo llvm-cov nextest (process-per-test), addressing intermittent Codecov “misses” on tracing::event! / log macro field-expression lines caused by cross-test callsite-interest caching.

Changes:

• Switch the CI coverage job to cargo llvm-cov nextest and install cargo-nextest in that job.
• Ensure coverage_attribute is enabled under coverage_nightly for non-test builds in crates that use #[coverage(off)] outside cfg(test).
• Clean up warnings uncovered by the new coverage build mode (remove redundant test imports; replace “uncoverable” no-op logging fallbacks with targeted lint allowances).

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
crates/fetch_hyper/src/lib.rs	Enable feature(coverage_attribute) whenever coverage_nightly is set so non-test coverage builds can compile #[coverage(off)].
crates/fetch_hyper/src/connection/hyper_handler.rs	Remove redundant test-only imports to keep coverage builds clean under -D warnings.
crates/cachet/src/telemetry/cache.rs	Remove non-executable no-op logging fallbacks and replace with cfg-gated allow(unused_variables) for non-logs/non-test builds.
crates/automation/src/lib.rs	Enable feature(coverage_attribute) under coverage_nightly (not just tests) to support coverage instrumentation builds.
Cargo.toml	Allow the unused_features lint workspace-wide to avoid coverage-instrumentation-only noise.
.github/workflows/main.yml	Run coverage via cargo llvm-cov nextest and install cargo-nextest in the coverage job.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[sandersaares]

This is smelly but I suppose there is no clean solution that avoids a cfg-conditional mess. We should file an issue in rust-lang/clippy to request possiblity for an allowlist in clippy.toml, though, so we can specifically say we are fine with "coverage_nightly" looking unused.

[psandana]

Approving, although I dislike the unused_features = "allow". We should fix this soon.

[sandersaares]

This is fixing the symptoms rather than the cause. If some test installs a global tracing listener (especially a no-op one) it needs to be isolated into its own test process via putting it into a separate test binary, thereby ensuring it does not contaminate state of other tests (alternatively, by modifying it to not install a global listener).

Hacking around this by using a different test runner is just that - a hack. This is not addressing the root cause.

This can cause other problems than just coverage gaps - it is a leak between tests, which is not something we should tolerate.

[ralfbiedert]

Would most of these go away once we switch to observed with individual sinks for tests?

[psandana]

Hacking around this by using a different test runner is just that - a hack. This is not addressing the root cause.

Shall we convert those from unit to integration tests ?

[sandersaares]

As I am on-call this week, I'll take over this topic as part of https://o365exchange.visualstudio.com/O365%20Core/_workitems/edit/7601362. Will try do the root-cause-addressing change in a separate PR, one that does not depend on tooling-imposed isolation.