Re-run and publish the Vistas 1024-parity and A40-control detections (#126, #137)

The Vistas resolution-parity result in #137 (docs/model_comparison.md, the 1024-input parity arm and the same-env 384 A40 control) rests on detections that no longer exist anywhere. The only copy was a private .model_cache shard in a scratch worktree on makelab2; a search on 2026-09-17 found no registered worktree under /homes/gws/jonf/RampNet, no shard written 2026-08-17..20, and none of the parity run's shard names (compare.cache_key over the 1024 signature) under /homes/gws/jonf, /tmp, /var, the root filesystem or the lab mounts. The doc states this beside the numbers; #137 was merged with the gap stated rather than held.

The fix is cheap on GPU — both A40 arms re-run in about 3m38s on makelab2 — but publishing the result into benchmark/model_detections/ runs into three things the roster's own rules leave open, and they need a decision before the re-run PR:

1. A pinned 1024 leg forces the published 384 file to be renamed too (test_every_leg_of_a_pinned_model_is_qualified), and the pin-naming rule expects a scalar.
2. The 384 control shares the published run's cache key by design, so it cannot be a pinned leg as the roster stands.
3. scripts/analysis/export_model_cache.py does not take --vistas-input-size.

Plan: settle 1–3 in rampnet/roster.py / the exporter, re-run both arms on makelab2, publish, and update the five doc sections that quote the parity numbers. The two tests that pin the pre-fix cascade state on purpose (tests/test_cascade_gate.py) will need the same treatment if the re-run does not reproduce the old numbers exactly — and if it does not, that is a result to record, not a discrepancy to smooth over.

🤖 Generated with Claude Code — Opus 5 (1M context), claude-opus-5[1m]

Start by reading rampnet/roster.py and scripts/analysis/export_model_cache.py, then inspect the relevant tests in tests/test_cascade_gate.py and the roster qualification tests. Run the existing tests before deciding how the pinned 1024 leg, 384 control, and input-size export should be handled. Done means both A40 arms are rerun, detections are published under benchmark/model_detections/, and the five documented parity sections reflect the recorded results.