Phase 6 full-scale dispatcher validation (141.6k A/B, ~$400) — gated on #684 merge

Goal

Run the production-scale Phase 6 validation of the multi-org API dispatcher: scripts/phase6_dispatcher_vs_baseline.py --n 141600 --confirm (the #658 workload shape, ~$400 total / ~$200 per arm). Same-session back-to-back A/B — legacy single-org judge_completions_batch vs the multi-org api_dispatch path, matched tier + model + N. Report wall-clock + cost SEPARATELY per arm (do NOT claim equal cost); confirm no data loss; target ≥5× wall-clock speedup.

HARD PRECONDITION — gate on #684

Do NOT run until task #684 has MERGED to main. #684 fixes the MAJOR bug where the dispatcher silently DROPS retryable items as terminal errors when a 429 hits on the final retry under a sustained storm — exactly what a 141.6k run against the shared rate limits triggers. Running before #684 lands wastes the $400 on a known-bad config and yields a misleading "data loss" result. If #684 ends blocked/archived instead of completed, do NOT run the test — surface to the user.

Cost authorization

User explicitly authorized this ~$400 spend ("run the 400$ test", 2026-06-27). Compute + log the projected cost from eval_results/api_throughput/phase0_limits.json prices before firing, then proceed (pre-authorized). If the projected cost exceeds ~$600 (>50% over the estimate), PAUSE and surface to the user instead of spending.

Constraints

• Shared org keys: obey §1b etiquette (per-key caps Sonnet100/Haiku120/Opus~40, AIMD back-off on 429, watch anthropic-ratelimit-*-remaining, fan across the 3 orgs).
• Judge model stays claude-sonnet-4-5-20250929.
• Save results to eval_results/api_throughput/phase6_full.json (+ per-arm metrics); update .claude/cache/api-throughput-status.md.
• On success: set-status <N> completed + post epm:progress with the headline numbers (per-arm wall-clock, per-arm cost, speedup, data-loss check). On a hard failure: set-status <N> blocked + epm:failure v1.
• Remind the user to ROTATE the 3 Anthropic API keys (still pending from #682).