Problem

The chat pipelines on both platforms contain the most bug-prone logic in the apps — streaming deltas, optimistic/provisional messages, durable history reconciliation, duplicate/out-of-order events — and none of it can be exercised deterministically today. Regressions in this area are found by hand on device.

Proposal

Add a deterministic, test-only replay harness per platform:

• Scripted gateway event sequences (streaming chunks, run completion, durable message arrival, duplicates, gaps, reconnect-mid-run) as fixtures.
• Tests drive the real chat view-model/controller with the scripted stream and assert the final transcript converges losslessly (no dropped, duplicated, or reordered content) and that intermediate UI state (pending runs, streaming text) behaves.
• Fixtures cover every markdown block shape so streaming rendering changes are also regression-guarded.

No production behavior change; iOS target is OpenClawChatViewModel, Android target is ChatController.

Impact

Streaming/reconciliation fixes stop regressing silently; future rendering and resume work lands against an executable spec.