[steph4n-gh]

The courier's strict-review path stops before a flagged move. Add a separate opt-in runner that requests an explicitly configured async expert immediately, then checks current context, permission and movement validity before returning its proposal. Attempts are capped per episode; callback failures, denied/invalid/stale/late answers and exhausted budgets stop it. Statistical review stays visible. Existing playground, core, Guard and email defaults are unchanged.

The equal-budget comparison selected the simpler policy. Progress tracking and plain bounded waiting produced identical per-map actions, completion, model calls, expert attempts and safety at every tested cap, so this PR adds no general progress-monitor API.

Independent confirmation: 40 paired-terrain maps	Completed	Actual fallback callbacks
Strict stop / cap 0	27/40	0
Each hybrid, cap 1 per map	35/40	13
Each hybrid, cap 8 or 256 per map	40/40	19

At cap eight, immediate added zero waits, progress 19 and plain bounded waiting 57; no unsafe proposal/dispatch or death was observed. Median elapsed episode times were 24.31 / 26.36 / 26.89 ms. These descriptive shared-host measurements include chosen 1 ms polling sleeps and overlapped full-suite execution; they establish no isolated speedup or customer latency savings. Direct known rules completed 40/40 with 1,410 cheap native invocations and no model calls, and remain the strongest inexpensive baseline.

This is the actual shipped courier execution loop with frozen saved skills, but its maps and teaching policy are synthetic. Fresh seeds 99300–99319 were frozen and scored once independently; the 40 worlds use 20 paired base layouts, and familiar role layouts/local inputs recur. It is not a paid or deployed real-world workflow or classifier qualification. The preliminary 98200 cohort is identified as selection/regression evidence. The committed protocol, source/skill hashes, compressed complete report and independent replay review retain all 560 confirmation comparisons.

Validation: 1,419 passed, 22 skipped in the full supported suite, including all eleven previously blocked localhost gRPC/HTTP cases; 39 focused tests and seven extra independent failure probes passed. Documentation and wheel/source archive checks passed. No model, corpus, core format or dependency changed.

Rollback: stop invoking examples.gaming.skill_playground.review_fallback, or omit its expert callback. StrictCourier(network) stops on review with no automatic fallback. No service configuration or saved skill is rewritten. One runner owns one episode/event loop; callbacks must be read-only and cooperative. Timeout discards late answers but cannot kill blocking code, and this example is no durable retry/restart/exactly-once ledger.

Draft for review only; no merge, deployment or release.