ci-fix: transient setup-task download failure on run 31619531465

[prql-bot]

Failure

tests failed on main in run 31619531465 (commit 1a82da5). Two jobs are red: test-prqlc-c / test-c and check-ok-to-merge.

Diagnosis

test-c failed while installing Task, not while compiling or running anything in this repo. The go-task/setup-task@v2 step (.github/workflows/test-prqlc-c.yaml:44) exhausted its three download attempts inside 24 seconds, all with the same transport-level error:

16:52:33.8548789Z socket hang up
16:52:33.8549282Z Waiting 12 seconds before trying again
16:52:46.0180073Z socket hang up
16:52:46.0180551Z Waiting 11 seconds before trying again
16:52:57.1854813Z ##[error]Failed to download version v3.52.0: Error: socket hang up

Everything before that step passed — cargo build --release --package prqlc-c, the minimal-c example, and the minimal-cpp example all completed, with the C++ example emitting correct SQL. Only task zig (which needs the Task binary) never ran.

check-ok-to-merge is the if: always() omnibus that lists test-prqlc-c in its needs: (.github/workflows/tests.yaml:678); its log shows every other dependency green. It is a consequence of the same failure, not a second one.

Why no fix was produced

The evidence points at ephemeral network infrastructure rather than anything in this repo:

• The same step passed hours earlier with identical config. Run 31601559466 (13:28 UTC) was green, and the commit under test changed only skill markdown — nothing that touches the C bindings or the workflow.
• socket hang up is a connection-reset at the HTTP layer, raised by the action's downloader against the GitHub release asset host. The release exists and resolves normally now.
• It is not recurring. This is the only failure conclusion across the last 25 tests runs on main, and no ci-fix: transient issue has been filed in the past 7 days.
• The version pin is already in place. The prior setup-task flake (ci: pin go-task version to avoid intermittent setup failures #5836, April) was a different shape — unable to get latest version from the default version: 3.x resolving through the releases API — and was fixed by pinning version: 3.52.0. That pin held here; the resolve succeeded and only the asset transfer failed, so raising retry counts would not have covered a window this narrow.

A GitHub incident affecting Issues, Pull Requests and Search (stspg.io/ssd9z8l2g46v) ran 16:16–16:41 UTC, ending 11 minutes before this failure and naming components unrelated to release downloads — noted for the record, but not offered as the cause.

Re-running the job should be sufficient. No code change is warranted.

[prql-bot]

Transient — closing as diagnosed.