Release: develop -> main

.config/nextest.toml

@@ -0,0 +1,48 @@
+# cargo-nextest configuration.
+#
+# Discovered from the workspace root (the directory holding the
+# top-level `[workspace]` `Cargo.toml`), so it applies to both
+# `cargo nextest run` (local runs) and `cargo llvm-cov nextest` (the
+# CI `Tests + Coverage Gate` job) — both drive the suite through
+# nextest and honour this file. It carries NO coverage semantics of
+# its own, so the 100% line/function gate is unaffected.
+
+[test-groups]
+# Concurrency cap for the heaviest Postgres-touching test module.
+#
+# Every test in `router::tests::jobs_endpoint_tests` runs
+# `crate::test_db::setup_pool()`, which CREATEs a fresh per-test schema
+# and replays the full migration suite (16 DDL files: tables, triggers,
+# views) into it. Postgres serialises concurrent DDL on shared system
+# catalogs, so when the post-#181 `--test-threads 8` default lets eight
+# of these migration replays run at once against the single shared
+# `postgres:17` container, each `setup_pool()` stretches from <1 s to
+# tens of seconds. The SSE `jobs_stream_*` tests additionally hold their
+# pool across deliberate `sleep`/`timeout` windows, so under that
+# contention their `JobStore`/`setup_pool` connection acquisition can
+# exceed the pool's 60 s `acquire_timeout` and surface as
+# `create: PoolTimedOut`.
+#
+# This was invisible at the `--test-threads 1` default. Capping the
+# group keeps the `Tests + Coverage Gate` (the single `ci:full` job,
+# `--test-threads 8`) safe: the `jobs_endpoint_tests` are interleaved
+# across ~440 tests there, but the cap still guarantees no more than
+# two migration replays race at once regardless of how the scheduler
+# packs the run. The cap originally surfaced under the now-removed
+# "DB Subset Tests" job, which selected `test(/^router::tests::jobs_/)`
+# alongside `db::tests` / `job_store::tests` etc. so the
+# migration-replaying tests clustered and the contention tipped over;
+# the cap is retained because it is the general guard for this group
+# in the full gate, not specific to that subset.
+#
+# Capping this group at 2 concurrent threads keeps useful parallelism
+# while bounding simultaneous migration replays so connection
+# acquisition stays well under the 60 s timeout. It touches no test
+# pool, so the deliberately-narrow error-path pools (`dead_pool`'s
+# `max_connections(1)` / 50 ms `acquire_timeout`) keep exercising their
+# `PoolTimedOut` arms verbatim.
+jobs-endpoint = { max-threads = 2 }
+
+[[profile.default.overrides]]
+filter = 'test(/^router::tests::jobs_endpoint_tests::/)'
+test-group = 'jobs-endpoint'

.github/workflows/auto-release-pr-staging.yaml

@@ -8,6 +8,7 @@ on:
 permissions:
   contents: read
   pull-requests: write
+  issues: write  # required by `gh label create`
 
 concurrency:
   group: auto-release-pr-staging
@@ -48,12 +49,6 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           COMMIT_COUNT: ${{ steps.check-diff.outputs.commit_count }}
         run: |
-          # Promote PRs intentionally do NOT apply the `ci:full` label.
-          # The heavy M3 Ultra test + coverage gate stays reserved for
-          # the develop → main Release PR (auto-release-pr.yaml), which
-          # remains the authoritative pre-PRD gate. Promote PRs run the
-          # slim `Lint & Build` job + Analyze / CodeQL, mirroring what
-          # every ready feature PR sees.
           printf '%s\n' \
             "## Automatic Promote PR" \
             "" \
@@ -64,14 +59,29 @@ jobs:
             "- [ ] Merge to promote staging to develop (deploys to DEV)" \
             > /tmp/pr-body.md
 
+          # `ci:full` opts the PR into the heavy M3 Ultra test +
+          # coverage gate (see ci.yaml). Promotions to `develop` deploy
+          # to DEV, so we want every promotion validated against the
+          # full gate (DB + prover + 100% coverage) rather than only
+          # the develop → main Release PR — apply the label on creation
+          # rather than relying on a human to remember the click.
+          # Mirrors auto-release-pr.yaml (develop → main).
+          gh label create ci:full \
+            --color FFA500 \
+            --description "Run heavy M3 Ultra test + coverage jobs on this PR" \
+            2>/dev/null || true
+
           # Created as DRAFT so the operator's `gh pr ready` is the
           # explicit gate that fires a `ready_for_review` event and
           # triggers ci.yaml — PRs opened via GITHUB_TOKEN would
           # otherwise hit GitHub's anti-recursion policy and skip
-          # downstream workflows entirely.
+          # downstream workflows entirely. The `ci:full` label is still
+          # applied at creation so the heavy gate runs as soon as the
+          # PR is marked ready.
           gh pr create \
             --draft \
             --base develop \
             --head staging \
             --title "Promote: staging -> develop" \
+            --label ci:full \
             --body-file /tmp/pr-body.md