Make the Bun delivery fallback test pass consistently #1232

Story

The delivery workflow’s Fable-to-Astra fallback test should give the same result each time it runs. It must verify that the workflow presents one hold when neither reviewer can complete.

Evidence

Bun shard 6/10 ran bun run test:bun 6/10 with Bun 1.4.0 on Linux at 92c27cd3c9ff69092a17ed536d24f39918ae9e7e. The Delivery Markdown test A-H required review follows Fable, one independent Astra replacement, then a retained Elicit hold failed:

Expected values to be strictly deep-equal:
2 !== 1
(fail) adversarial implementation example > runs the Delivery Markdown journeys without native models

The full first log is linked above (SHA-256 661e9bb04b920384279e6f8fa0a82f862fd9e736e97f806d3e9f5500f4801052). It does not identify which count assertion failed.

The same revision passed the focused journey unchanged on macOS arm64 with Bun 1.3.14: one test passed, zero failed, 168.47 seconds total.

bun test --timeout=300000 packages/cli/tests/adversarial-delivery.test.ts

The preceding Linux Bun shard passed with the same fixture source. The intervening change was in ACP load diagnostics. This comparison does not establish the cause of the count mismatch.

The failed-jobs-only CI retry passed the same Linux Bun shard on the identical commit. No code changed between these two CI attempts.

Acceptance

Identify the specific count and conditions that let it differ. Keep the one-review and one-hold assertions, and show that the journey passes repeatedly under the affected Linux Bun CI conditions.