Render Python Stats

Aggregate python-mode render stats over the last days days.

Returns a snapshot good enough for ops to decide rollout cadence:

• attempt distribution (1/2/3) → MiniMax stability proxy
• failed_rate → how often we exit without a usable pptx
• avg_attempts → cost proxy (higher = more retries = more MiniMax calls)
• hardest_slides → top N slides with attempts ≥ 3 OR errored, including title + project + error so you can drill into specific failures

Caveats:

• "deterministic_rate" can't be derived precisely from current Firestore fields. attempts==3 + status ok could be either "MiniMax 3rd attempt succeeded" or "deterministic fallback kicked in". To distinguish we'd need to start writing a python_render_strategy field; not yet done. For now, attempts==3 is reported as a "needs-investigation" bucket.

Auth: ADMIN_USER_IDS allowlist (env). Non-admin → 403.

Query is a single Firestore range scan on slides where render_mode_used == "python" and first_render_charged_at >= cutoff. Per slide, we pull only the fields used in aggregation — kept O(N) where N is rendered-slide count over days. At current traffic this is well under 1k docs / 7 days, so loading them all in memory is fine.