test(e2e): retry flaky tests with cargo-nextest

.config/nextest.toml

@@ -0,0 +1,17 @@
+# Nextest is used only by the e2e harness (crates/e2e/mise.toml `run` task,
+# always with `--profile e2e`); the hermetic unit suite stays on plain
+# `cargo test`, which compile-checks feature-gated targets in a way nextest
+# does not replace.
+
+[profile.e2e]
+# Retry a failed test up to 2 more times before counting it as failed. The e2e
+# suite occasionally flakes on GitHub runners (kind/apiserver hiccups), and a
+# failed required check evicts the PR from the merge queue with no automatic
+# requeue — so absorb the transient failure inside the check run instead.
+# A retried-then-passed test is reported as FLAKY, keeping the flake visible
+# in the job log rather than silently swallowed.
+retries = 2
+# The suite is stateful against one shared kind cluster and was previously run
+# with libtest `--test-threads=1`; keep it serial. (`--no-capture` on the CLI
+# also forces this, but state it here so the profile stands on its own.)
+test-threads = 1

.mise/config.toml

@@ -36,6 +36,9 @@ kind = "0.32.0"
 kopia = "v0.23.1"
 "aqua:EmbarkStudios/cargo-deny" = "0.20.2"
 "aqua:taiki-e/cargo-llvm-cov" = "0.8.7"
+# Runs the e2e suite (crates/e2e `run` task) so individual flaky tests retry
+# in-process-tree instead of failing the whole CI shard; see .config/nextest.toml.
+"aqua:nextest-rs/nextest/cargo-nextest" = "0.9.140"
 # Docs site toolchain. The docs are an MkDocs Material site (see `mise run docs`);
 # the MkDocs + Material + pymdown-extensions versions are pinned in pyproject.toml
 # / uv.lock and run via uv, so the only mise-managed tool here is uv itself (which

crates/e2e/mise.toml

@@ -356,20 +356,25 @@ set -eu
 mkdir -p target/e2e; rm -f target/e2e/.diag-done
 trap 'rc=$?; if [ "$rc" -ne 0 ]; then echo "==> e2e tests FAILED (rc=$rc) — dumping cluster diagnostics" >&2; mise run //crates/e2e:diagnostics || true; touch target/e2e/.diag-done; fi; exit $rc' EXIT
 # KOPIUR_E2E_BINS — space-separated test-binary (file) names, expanded to repeated
-# `--test <name>` flags BEFORE `--`, so a CI shard runs only its files. Empty ⇒ all.
-# KOPIUR_E2E_TESTFILTER — libtest name-substring filter, kept AFTER `--`.
+# `--test <name>` target flags, so a CI shard runs only its files. Empty ⇒ all.
+# KOPIUR_E2E_TESTFILTER — name-substring filter (nextest positional filter).
 BINS=""
 for b in ${KOPIUR_E2E_BINS:-}; do BINS="$BINS --test $b"; done
 # The CLI e2e (tests/cli.rs) runs the compiled plugin binary (`kopiur`) as a
 # subprocess; build it here so the tests never race a stale/missing binary.
 echo "==> building kopiur CLI (e2e subject)"
 cargo build -p kopiur-cli --locked
 echo "==> running e2e tests${KOPIUR_E2E_BINS:+ (binaries:$KOPIUR_E2E_BINS)}"
-# `--no-fail-fast`: run EVERY test binary even when one fails, so a single flaky bin
+# nextest instead of `cargo test` for the `e2e` profile's per-test retries
+# (see .config/nextest.toml): a transient failure retries in place instead of
+# failing the CI shard and evicting the PR from the merge queue.
+# `--no-fail-fast`: run EVERY test even when one fails, so a single flaky bin
 # (e.g. the heavyweight NFS/server backend test) can't hide the pass/fail of the rest.
-# cargo still exits non-zero overall, so the trap + CI failure are unaffected.
-cargo test -p kopiur-e2e --features e2e $BINS --no-fail-fast -- \
-  --include-ignored --test-threads=1 --nocapture ${KOPIUR_E2E_TESTFILTER:-}
+# nextest still exits non-zero overall, so the trap + CI failure are unaffected.
+# `--run-ignored all` ≙ libtest `--include-ignored`; `--no-capture` streams
+# output live and keeps execution serial, like the old `--test-threads=1 --nocapture`.
+cargo nextest run --profile e2e -p kopiur-e2e --features e2e $BINS --no-fail-fast \
+  --run-ignored all --no-capture ${KOPIUR_E2E_TESTFILTER:-}
 echo "==> e2e tests passed"
 '''