[rmcdaniel]

mirror.gcr.io has twice failed to serve one blob for a few minutes, with docker: failed to copy: httpReadSeeker: failed open: could not fetch content descriptor …:

• a MariaDB 10.11 layer (3e968919…), 00:42–00:47 UTC (Rust #325: note takeover's planned deprecation in 3.1.x #930 PR run, then Rust #325: close a run whose deadline passed when a task failure is reported #926's merge group 41367763841);
• the MySQL 8.4 image's in-toto attestation (dfe2bc17…), 03:01 UTC (Rust #325: close a run whose deadline passed when a task failure is reported #926's merge group 41379804081).

Each time, a deadline-health cell's plain docker run failed, and the merge queue dropped #926. The image pulled fine a few minutes later.

Change

codec-operations-cell.sh pulls its runtime, observer and database images before it starts anything, with up to four attempts at 15, 30 and 45 s apart, and a warning annotation for each failed attempt. All three images are pinned by digest, so a repeat pull fetches the same content. A pull that fails four times still fails the cell, as before.

Checks

• node --test tests/Unit/ServerParityCodecDeadlineCellsTest.mjs: 9 pass. The test lifts the cell's cleanup() only, which is unchanged.
• bash -n on the script.
• The function, extracted and run against a fake docker, succeeds on the third attempt after two failures and refuses after four.

🤖 Generated with Claude Code

[rmcdaniel]

Security bounded review of exact head e3be66eabd6dc1c4ff9168f163a53f4fbd2a9d76: clear; no Security hold. The helper retries exactly four times, delays only between attempts (15/30/45 s), and returns failure after the fourth so set -e still fails the cell closed. All three possible inputs are digest-pinned at this head (runtime, observer, and every non-SQLite database image); the repeated pull cannot float tags. Pulls happen before the network or containers are created, so an exhausted retry needs no earlier cleanup trap. The annotation contains only the public image reference. I independently checked script syntax and the committed pin formats; the owner-provided fake-Docker success-on-third/fail-on-fourth controls are proportionate. This CI-only hardening does not alter workload retries, deadlines, product behavior, or rc.3 authority.

[rmcdaniel]

General bounded source review of exact e3be66eabd6dc1c4ff9168f163a53f4fbd2a9d76: clear. The sole 19-line script delta pre-pulls the unchanged digest-pinned runtime, observer and selected database images before creating any network/container. Exact manifest references are pinned at this head. Each failure is visible, waits are 15/30/45 seconds, success returns immediately, and the fourth failure returns nonzero under existing set -e; there is no tag/registry fallback or swallowed terminal failure. SQLite skips the database pull. Existing cleanup, cell resources, conformance assertions and job-level limits are unchanged.

This is source review only. I did not execute native fixtures, pull images, run a benchmark or resume stopped observation qualification. The reported fake-Docker success/exhaustion controls and 9 Node tests are owner-reported. No rc.3 gate/order change requested.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud