Collect the versions benchmark results hourly

[alexey-milovidov]

Adds .github/workflows/versions-collect-results.yml — an hourly job that looks at the sink for new versions benchmark results, mirroring collect-results.yml for the main benchmark.

The difference is that the versions benchmark has no materialized view: a machine POSTs its whole result JSON to sink.data with kind: versions-benchmark, and versions/fetch-results.sh reads the results back out of those raw rows. The new versions/collect-new-results.sh is the automation around it:

1. fetch the results that arrived in the last day (SINCE_HOURS, dispatch input full: true for every version),
2. rebuild versions/data.generated.js when the result files changed,
3. commit both to auto-results/versions, open a pull request and merge it — these are ClickHouse's own results, trusted the way the main collector trusts the clickhouse* systems,
4. report runs that sent a log but no complete result (still running, crashed, or an incomplete load) in the job summary — nothing else would show those.

Supporting changes:

• fetch-results.sh gains SINCE_HOURS=<n>: fetch only the versions that ran in that window and leave the other files alone, instead of refetching all ~200 versions from the sink. That is what makes the job cheap enough to run hourly. A full refresh is still the default when the variable is unset, and the collector refuses to publish one that loses more than a tenth of the result files — that means a broken query, not lost results.
• The collector restores result files that come back byte-different but identical in content. jq reformats numbers (0.020 → 0.02) whenever it reconstructs a document, so apply-minvers.py alone would otherwise commit a few hundred formatting-only diffs every hour.
• The versions README's claim that a materialized view builds the report is corrected, and the collection path is documented.

Verified locally against a stubbed clickhouse-client in a scratch worktree: with no new results the tree is left completely clean (200 re-serialised files restored, nothing committed, data.generated.js not even regenerated); with one new result only that file and data.generated.js change and the title/body come out right; a full refresh that loses results exits 1 without publishing; generate-results.sh is byte-stable across runs. Not exercised: the real database queries and the PR/merge step, which need the CLICKBENCH_DB_PASSWORD secret and run only in Actions.

🤖 Generated with Claude Code