Complete downloads and transcript write limits

New releases remain drafts until all three platform archives and their checksums pass download verification. The build uses the release commit, and the published downloads must match the build files. A failed build or download check leaves the previous complete release available. PR #1157 also makes retries verify existing published files without replacing them.

Large Claude, Codex, Pi, and OMP transcripts can fill the write buffer before a file ends. PR #1158 loads rows when the buffer reaches 25,000 rows or 32 MiB of serialized UTF-8 data. A complete incoming row can take the buffer above the byte limit. These limits apply to the buffer, not total process memory.

Periodic progress marks let interrupted runs skip completed files on retry. The failure tests check concurrent writes, interruption, and recovery after row or watermark writes fail. A Codex test compares stored records with direct database writes. It also verifies that rows reach the database before file processing ends.

PR #1155 corrects the setup instructions. Run the first import explicitly:

ax update
ax --version
ax ingest

Later read commands can refresh stale data. Improvement task instructions now work with the user's selected coding agent.

The release range records these changes. The release notes use the code changes and direct test results as evidence.

Generated with ax.