Make trainer catch up on v1 self-play data

[k15z]

Summary

This changes the trainer startup/resume path for the existing v1 model + accumulated self-play data case.

• Train v1 catch-up against the full self-play corpus until v2 exists, instead of applying the sublinear replay window immediately.
• Use a v1 catch-up token bucket sized for catchup_passes_over_existing_selfplay passes over existing self-play, without the normal one-report seed/token caps.
• Force the v2 candidate export once catch-up passes are complete, while deferring ordinary fresh-position candidate exports during full-corpus mode.
• After v2, return to the normal rolling replay window and token bucket behavior.
• Persist a resumable trainer checkpoint under state/trainer/resume_checkpoint.pt with model weights, optimizer state, train bucket state, SWA buffer, BN refresh batches, and catch-up counters.
• Rename trainer-side model turnover helpers from promotion language to candidate publishing. Evaluation service promotion/approval is unchanged.
• Reduce promote_every_new_positions from 2.5M to 1M for normal candidate exports after v1 catch-up.

Why

We already have a v1 model and around 1.1M paid-for self-play samples. The old restart behavior could apply the sliding replay window and normal token-bucket seed cap immediately, meaning a restarted trainer could ignore a large share of existing v1 self-play before ever producing v2. This PR makes v1 catch-up explicit: deploy it, train across all existing self-play, publish v2, then continue v3/v4/etc. on the normal rolling cadence.

The trainer summary feature is intentionally retained as observability/reporting cadence. It is not used as the candidate boundary.

Deployment notes

• Existing v1 latest.meta.json without positions_at_promote is treated as watermark 0 so legacy v1 artifacts can start catch-up.
• For v2+, missing positions_at_promote still fails because the watermark should be explicit.
• models/checkpoint.pt remains a last-published-model alias, not a live trainer snapshot. The new resume checkpoint is state/trainer/resume_checkpoint.pt.

Validation

• uv run pytest -q -> 186 passed, 1 skipped
• git diff --check
• codex review --uncommitted -> no discrete correctness issues