Final column artifacts and the scope of the older gates

This narrow qualification successor is based on the exact PR18 candidate 35b00247e68f04e6db539c365298862f82e1c2d0. It changes no Rust physics or native outputs. Historical receipts, examples, figures, Lean files and published book inputs remain frozen. External review and PR advancement stay with the parent.

What the old gate omitted and what is repaired

The column-interface companion promised finite fields, accepted geometry and native pixels. Its reader checked initial/frame exports and final cells, but never read coupled final-faces.csv, final-geometry.csv or final.png. The old plot generator read two of those unchecked artifacts to display the accepted eight-interval result. Fresh frozen-reader controls therefore accepted nonfinite final velocity/geometry and a corrupt final PNG in normal and optimized Python. This was a reader coverage defect; the stored original artifacts remain valid and are not replaced by the corrupt witnesses.

The successor parses every coupled final face/geometry field through the same explicit finite-field checks, verifies exact identity with the last accepted frame, and rederives every final grayscale pixel from that frame’s liquid fractions. Integer face indices are required before conversion to dictionary keys, preventing fractional indices from aliasing valid faces. Missing finals, finite changed fields, nonfinite fields and malformed/wrong-pixel PNGs reject. Imposed fixtures export no final face/geometry/PNG triplet: their single imposed face field and final fractions retain their separate manufactured oracle.

New controls exercise ten corrupt-final variants on all six coupled scenarios, in both Python modes, alongside the twenty original controls. The native thirteen-scenario packet and sixty-four coupled intervals must still pass. The guarded successor plotting run copies immutable inputs into temporary storage, verifies that copy, regenerates the independent summary and invokes the frozen plot code there. It saves the new figure separately and never writes over the historical figure or receipts.

CI changes

Both pull-request and main-push path lists cover examples/**, replacing the incomplete explicit example list. They also cover both column-interface reader files and evidence/column-interface/**. A direct test checks each of the candidate’s twenty-three example paths, both verifier paths and a final native PNG as individually changed files for each event. This closes the reported eleven-example omission and also covers future examples. It changes triggering, not solver settings or the CI job’s existing checks.

The parent reports that CI run 37424773368 ran from 06:38:53 to 06:54:12 and ended cancelled. Formatting and all strict lints passed, 237 default and 231 core-only tests completed, desktop dependencies compiled for 3m16s, and 85 desktop tests ran before cancellation. Smoke export and lifecycle checks were skipped. The log says operation cancelled; exhaustion of the 15-minute job budget is a supported diagnosis, not a confirmed terminal timeout label. Direct API retrieval is forbidden in this environment, so these run facts are preserved with their parent-reported provenance and the frozen workflow source.

The smallest configuration repair retains the shared serial job/cache and raises its budget to thirty minutes. The original serial compilation/testing already exhausted fifteen minutes before completion of desktop tests and the separate release CLI build/smoke/lifecycle stages. Thirty minutes reserves another old job budget for the remaining builds/checks and ordinary runtime variation. A byte comparison of the whole job after substituting the old timeout proves that no feature mode, command, assertion or evidence-upload step was removed. Local remaining-stage runs are measured separately; they are warm-cache execution, not evidence that a fresh hosted runner will finish in the same time. A new hosted run must complete every stage before CI qualification can be claimed.

Limits this repair does not close

tools/verify_viscosity.py validates isolated manufactured modal behavior and the coupled demonstration’s finite fields, energy, phase ledger, divergence, versions and pixels. After the first symmetric force pulse it does not independently replay coupled vector transitions through advection, forcing, viscosity and pressure. The coupled energy comparison is consequently not a complete dynamics oracle: a same-energy divergence-free field change can evade those checks if corresponding pixels are changed. The existing isolated viscosity tests do not supply that missing coupled transition check.

The column-interface reader independently checks the initial pulse/activation pressure and all mixed-rest pressures, later held-wet divergence, atmospheric air storage and held/end geometry. It does not reconstruct the later changing wet pressure solve or pressure/velocity correction from the prior frame. A finite fabricated later wet pressure can therefore evade the gate. This fails to cover an inference that all later pressure values are numerically correct; classification activation and accepted publication are narrower evidence.

These limits remain explicit scope witnesses, not positive physical qualification. The newer coupled-discrete/third-component replay applies to a different declared model and does not repair either older reader. No new Lean, moving-cap pressure accuracy, general liquid or continuum validation claim is made by this artifact/CI repair.