End-to-end wall-clock table + checkpoint archive with hashes

[duckyquang]

Two bounded items, both yours since you own the wall-clock work:

1. A deployment-facing wall-clock table for the full paper: for the repair pipeline (nonlinear R10 + the R28 arms), per-arm total wall time per instance measured at the same boundary (your Per-arm wall-clock is not measured at the same boundary across arms #104 fix), on one machine, one table - ranker forward + 1 solve vs probe vs enumeration vs restart curve. The paper claims cost wins; this is the receipt. New script or extend run_repair_ranker --solve-e2e output; do not change any protocol.

2. Checkpoint archive: every citable checkpoint (repair_random_support_holdout, repair_nonlinear_balanced_full, geo_repair v3 when it lands) tar'd with a SHA-256 manifest committed to PROVENANCE, upload location per lab convention. Closes the artifact reviewer's "regenerable is not the same as available" flag.

Est 1.5 days.