Add comprehensive rendering regression tests for feed ownership and streaming behavior #183

The feed renderer needs proper regression coverage. We have repeatedly reintroduced bugs around duplicated assistant messages, missing user prompts, stuck live turns, semantic history handoff, ghost fallback, and live vs committed tool rendering. These bugs come from multiple channels trying to own the same visible feed content: committed transcript entries, semantic current turns, semantic history, ghosts, optimistic user rows, and queued prompts.

Add a focused rendering test suite that protects the core invariant: each visible feed artifact should have exactly one owner at a time, and ownership should transfer cleanly when committed transcript data catches up. The tests should cover the critical Claude and Codex rendering paths without requiring live providers.

This should include the known high-risk areas: committed-vs-semantic duplicate suppression, semantic history bridging JSONL lag, ghost fallback rules, optimistic/queued Codex prompt behavior, stale live turn recovery, and custom tool rendering/streaming for important tool calls like edits, writes, apply_patch, exec_command, and write_stdin.

Acceptance criteria:

• Rendering regression tests run through the standard test suite or a documented focused command.
• Tests cover the known duplicate assistant response and missing user prompt classes.
• Tests cover semantic current/history/ghost ownership interactions.
• Tests cover Codex optimistic submit and follow-up prompt queue behavior.
• Tests cover at least the most important live tool rendering paths.
• Tests use explicit assertions over ownership/render output rather than broad brittle snapshots.
• Future rendering changes should fail tests if they reintroduce duplicated rows, missing rows, or stuck streaming state.