mu-9bri/mu-b20l solo: chronological queued-prompt commits + tracked finalize (no more stutter)

[tcovert-c137]

Two transcript-fidelity fixes from live operation (sessions 64dcebc762f83eb9), fixing mu-b20l and mu-9bri.

mu-b20l — assistant text duplicated around every tool call ("the stutter")

Every text preceding a tool call rendered twice: above the ● marker and again between it and the ⎿ result. The event log shows one canonical text block per message — pure render artifact. The wire orders text_delta… → tool_call_started → assistant_text_finalized, so at finalize time the trailing item is the ToolCall; the old last-item-only check appended the canonical text instead of replacing the streamed item.

Turn now tracks the in-flight streamed item by index (streaming_text_ix / streaming_thinking_ix): deltas extend the tracked item, finalize swaps it for the canonical text and ends tracking, and a finalize whose invocation streamed nothing pushes fresh instead of clobbering an earlier invocation's already-canonical item. Applied to both text and thinking channels; order-independent.

mu-9bri — queued prompts commit in chronological position

A prompt queued mid-turn committed below the whole finished turn, so the part of the turn that answered it rendered above the question (observed with a prompt typed during a blocking dialogue_poll). PendingInterjection now captures the live turn's item count at queue time, and the commit interleaves per plan_splice_commit: items before the splice → the prompt (keeping its queued-while-responding label) → the items that answered it.

Deliberate granularity: the item boundary is authorship-correct, not an approximation — deltas still streaming into the in-flight item come from a provider call that never saw the prompt (absorption injects it at the next InvokeLlm), so the answer necessarily begins in a later item. Bridge-gap prompts (no live turn at queue time) resolve to the next spliced prompt's position when one follows — a panel round-3 finding, fixed and pinned in tests — and to the legacy after-the-turn position otherwise. Turns with no spliced prompts keep the single-block commit path byte-for-byte. Settlement bookkeeping (mu-z9ol receipts) is unchanged.

Verification

• 153 mu-solo tests green, including new coverage: tracked-finalize across trailing tool calls, no-delta finalize non-clobber, splice-plan chronology and edges (start/past-end/same-point/mixed bridge+mid-turn both orders).
• Full just ci green.

Gate

just ci-aipr: first panel escalated with gpt-5.5 holding a mixed-case planner objection; the objection was correct, got fixed (the [None, Some(k)] resolution above) and pinned in tests, and the re-gate converged unanimous approve in 1 round.

🤖 Generated with Claude Code

https://claude.ai/code/session_01REB3QZvP3arA2cLMVRgYam