Add clean-replay MapGPT and SpatialVLA tasks
Selection note
This PR contains the intersection of completed, comparable evaluations where Opus-5 outscored hy3 and the task has the clean-replay check revisions applied.
| task | Opus-5 | hy3 |
|---|---|---|
| mapgpt-r2r | 0.530340 | 0.380042 |
| spatialvla-libero | 0.565383 | 0.491082 |
pi06-libero-recap is intentionally excluded: its hy3 0.000 run has no rubric_results.json, so it is not a comparable completed-evaluation artifact. pi05-libero is not part of this selection.
Both packages use a clean low-privilege replay, immutable frozen candidate evidence, source-manifest checks, and evaluator-only scoring against EVIDENCE_ROOT. The public MapGPT package omits verifier-private goal paths.
Superseded by #8, which carries all 15 robotics task packages at the 2026-09-06 state plus verifier_refs_required.json. This branch predates this week's material fixes, and five partially overlapping robotics PRs made it impossible to tell which version was authoritative -- the same class of confusion that produced the image/package mismatch found by the 2026-09-04 runtime audit. Closing in favour of #8.