refactor(ai): drop responses replay tombstones

[rekram1-node]

Why

ParserState carried two tombstone sets, completedMessages (#46064) and completedTools (#45789), that ignored output_item.added / output_item.done for item ids the parser had already finished. #46064 justified this with an OpenAI background-mode stream resumed via starting_after, where the resumed stream can replay events for already-completed items.

Our client never sends background: true or starting_after on Responses requests and never tracks sequence_number; parser state is built fresh per attempt via initial(request), and a disconnect is a new request rather than a resumed stream. The WebSocket channel continues between turns with previous_response_id, not mid-stream. So the replayed-event inputs these guards handle cannot reach the parser. The spec side agrees: a completed item receives no further updates and output_item.added is emitted once per item.

Neither Codex nor pi tracks completed ids or defends against replayed item events, for the same reason — neither resumes streams.

What changes

• Remove completedMessages and completedTools from ParserState and initial, plus the branches that consulted or populated them.
• onOutputItemDone for a message no longer early-returns when a different message is active; it finalizes its own item and clears state.message only if that item was the active one.
• Kept: done-only messages and function calls still emit a full lifecycle from the terminal item (that part of fix(ai): preserve done-only response messages #46064 removed a condition rather than adding one, and matches how both peers treat output_item.done as authoritative). Duplicate added for a still-pending function call is still a no-op via state.tools. completedCompactions stays because response.completed.output legitimately re-lists compaction items during terminal recovery. Reasoning dedup via reasoningItems is unchanged.
• Tests: removed the four tests whose fixtures were replayed or reused item ids; stripped duplicate done events from two fixtures whose remaining assertions still hold.

Verification

cd packages/ai
bun typecheck
bun run test   # 976 pass, 28 skip, 0 fail