What

• derive bridge routing only from turn/provider/profile data; no backend or model fallback
• reuse caller execution IDs or deterministic wire-safe IDs
• continue sessions and replay from server event cursors
• cancel active bridge jobs on reader exit or environment shutdown, waiting for confirmed server completion
• reconstruct full replay results from the bridge-owned run instead of keeping a second local history

Why

cli-bridge jobs survive client disconnects. The SDK provider previously neither sent stable run IDs nor cancelled those durable jobs, and it silently selected OpenCode when no caller selected anything. That made resume unreliable and shutdown capable of leaking paid work.

Proof

• clean workspace test run: 697/697 unique tests passed
• provider tests: 29/29 passed
• workspace type checks and builds passed on the exact commit
• isolated package check passed: 8 packages, 18 exports, 36 export targets
• live route check: Pi 0.75.5 exposed pi/tangle-router/glm-5.2; the profile-driven smoke returned exactly READY in 16.4 seconds (15,053 input / 3 output tokens)
• independent adversarial review: approved exact diff; 0 named provider/model/backend cases

Remaining upstream issue

cli-bridge must serialize concurrent turns that share one backend/session pair. That enforcement belongs in cli-bridge across all clients, not in this provider.