Improve training observability and next-run defaults

[k15z]

Summary

• add an April 2026 postmortem documenting the failed v1->v5 run and the evidence behind the main failure hypotheses
• persist immutable trainer export metadata and version-scoped promotion summaries, and keep trainer JSONL logs durable in S3
• switch the next run to 1200-sim / 10x192 / full-search-only self-play, keep bootstrap imitation, but remove imitation from replay after bootstrap
• make evaluation collection fixed-budget so candidate comparisons stay comparable even after an early provisional SPRT result
• update tests that were pinned to the old defaults

Details

• write models/versions/{N}.meta.json and state/runs/v{N}/promotion_summary.json for each export
• include metadata links in latest.meta.json and approved.meta.json
• continue collecting gate and benchmark games until the fixed budgets are complete before final promote/reject decisions
• add trainer artifact tests for immutable metadata and promotion summaries
• document the April run in postmortems/POSTMORTEM_APRIL_2026.md

Validation

• uv run pytest training
• cargo test
• git diff --check