G1 — Chat-stream/main feasibility study (design doc, no production code)

Produce the design gate G1 study required by plans/cleanup-state-machines.md ("Design gate G1 — Chat-stream/main feasibility study"). Read that section for the ten questions and five required artifacts (serializability inventory, target state/read-model schemas, acceptance transaction, migration sequence, deletion budget).

Inputs to read: the G1a DECIDED section (status/error are settled — do not relitigate); recorded product decisions 1–5 and the Remote intent policy; the actor lifecycle matrix row for chat_stream; plans/distrbuted-machines.md ("Chat-stream feasibility study" section and the durable protocol actor pattern); plans/claude-cleanup-machines.md recipes for chatMessagesByIdAtom, the queue pair, and planStateAtom (reader inventories and invariants to preserve — recipes to evaluate, not preapproved); src/chat_stream/ and src/user_input/follow_up_handoff.ts as they exist today (the memory-only handoff design and its documented structural-safety argument are inputs to the durable-acceptance question).

Deliverable: plans/g1-chat-stream-study.md — decision-forcing: each question gets a recommendation, the alternatives, and rejection reasons; ends with (a) a go/no-go recommendation for the C3 host move, (b) the A6b subset that is safe to build renderer-side regardless of the host outcome, and (c) the updated chat_stream row of the actor lifecycle matrix. Where a recommendation depends on unmeasured performance (chunk fan-out, snapshot frequency), name the measurement and its threshold instead of guessing.