ci(quality): retry ollama model pull on transient network failure

[planetf1]

Fixes #1233

What

The Pull models step in .github/workflows/quality.yml calls ollama pull granite4.1:3b with no retry logic. A single connection reset by peer from Ollama's CDN fails all three Python matrix jobs and requires a manual re-run.

Wraps the pull in a retry loop: 5 attempts, 20-second backoff, ~2 minutes of total headroom — enough to ride out transient CDN blips while still failing fast on real errors (wrong model name, Ollama not running, sustained outage).

Why not cache?

Caching the model blobs is a follow-on improvement. The retry loop addresses the flakiness with a two-line change; caching requires knowing the correct system path for Ollama's model store on the GitHub-hosted runner and managing cache invalidation. Filed separately.

Testing

Observed the failure mode on PR #1174 (connection reset by peer on granite4.1:3b). Re-run passed. The retry loop was added and the subsequent CI run for #1174 is currently in progress with the fix in place.

[akihikokuroda]

LGTM

[markstur]

Seems like on the 5, it will echo "retrying..." and continue successfully (to whatever error happens later), when it probably should instead exit 1 on that last attempt.

Something like this (untested):

for i in 1 2 3 4 5; do
  ollama pull granite4.1:3b && break
  if [ $i -lt 5 ]; then
    echo "Attempt $i failed, retrying in 20s..."
    sleep 20
  else
    echo "Attempt $i failed, no more retries"
    exit 1
  fi
done

This fixes:

1. Exits with code 1 if all attempts fail
2. Shows meaningful message on final attempt instead of retrying...

But I'm not going to -1 this because I'm sure what you have generally works and solves the problem well enough. Also, looking forward to future caching!