fix: stop the refresh loop faking convergence, and survive a failed publish

[gabrielspadon]

Found while diagnosing why yesterday's refresh failed. The 403 was the visible problem; the convergence bug underneath it was worse, and it was mine.

The refresh has never actually converged

Removing the response cache deleted if ! save_cache "$i"; then but left its body and its fi — and that fi was the one closing the pipeline-failure branch above:

if [ "$rc" -ne 0 ]; then
  echo "::error::Pipeline run $i failed..."
  ...                                  # ← everything below is now INSIDE this branch
  digest=$(find output ... | sha256sum ...)
  echo "::error::Cache save failed after successful pipeline run $i"
  exit 1
fi                                     # ← the orphaned fi, ~25 lines late

if/fi counts still balanced 4-to-4, so nothing caught it. But on a successful run that whole block is skipped, so digest and prev_digest are both unset and compare equal:

if [ "$digest" = "$prev_digest" ]; then   # "" = ""  → true, on iteration 1
  echo "Corpus unchanged across two consecutive runs — converged"

Run 32033375617 shows it exactly: Pipeline run 1/10, runs_completed=1, "converged" — and no Run N — corpus digest … line at all, because that echo was in the dead branch too. One pass, declared as two. The 403 accidentally prevented publishing an unverified corpus.

Fixed by closing the block where it belongs, rejecting an empty digest, and running the step under set -uo pipefail — which is what would have caught this originally.

Verified by simulating the loop with a stub pipeline: a corpus that changes for three iterations now converges at run 4, not run 1.

Publication no longer discards an hour of quota

The 403 destroyed a completed corpus because three things lined up: the push had no retry, output/ lives only on the runner, and the verify job neither re-triggered nor preserved — the re-trigger path required result == 'success', and the failure path just exit 1.

• pushes and PR calls retry 6× over ~8 min (diagnostics to stderr, since one call sits inside $(…))
• a failed push uploads output/ as an artifact
• a converged-but-unpublished run re-triggers; only a run that never converged is a failure

Four contracts, each mutation-tested

Mutation	Result
re-break the fi pairing	CAUGHT
drop the empty-digest guard	CAUGHT
remove retry from the push	CAUGHT
remove the publish-failure re-trigger	CAUGHT

bash -n over every run block is the one that matters: it sees mispaired blocks that a count cannot.

Verification

ruff clean, mypy clean (55 files), pyrefly 0 errors, pytest 2021 passed, 2 skipped. Every workflow block shellchecks clean at -S warning.

Note: GitHub is still mid-incident (degraded API/Actions), so CI here may be flaky for reasons unrelated to the diff.

[qodo-code-review]

ⓘ Qodo reviews are paused because the subscription is no longer active. Ask your workspace admin to reactivate the subscription to resume reviews. Manage billing