fix(core): keep between-tool narration brief - #59035

The prompt edits are consistent and the affected tests pass (session-system-prompt, plugin/optimize, session-runner-recorded). I couldn't check the actual effect on model behaviour because there are no provider credentials and the simulated model doesn't follow prompts. Public API is unchanged: only built-in prompt text and a recorded test fixture change.

The main concern is the premise. The new instruction tells every model that the user doesn't read text between tool calls, but in v2 that text is shown in the TUI and app. As the description itself says, "the text still renders in v2". On #59012 the bot noted that some users rely on it to follow progress. Telling the model something false about its UI to change its style is fragile, and making that text unread by default is a product decision for maintainers. The issue is a feature request, and users can already opt in through AGENTS.md/instructions.

If the direction is accepted:

• The new line contradicts existing text in meta.txt and trinity.txt (inline).
• The GPT rewrite drops the commentary guidance: stating assumptions and flagging tradeoffs and blockers mid-task (inline).
• "make a sensible choice and keep working" with no mid-task questions may push models to act on their own where they would otherwise check first.
• The new test only checks that one exact sentence is present in each prompt file, which locks in the wording and checks no behaviour. I'd drop it to keep the change minimal.

Related: #55298

Both earlier points are addressed. The new line no longer contradicts "all text you output outside of tool use is displayed to the user" in meta.txt/trinity.txt. The GPT commentary guidance is back to the base version. The new wording only asks for brief text between tool calls and moves questions and outcomes to the final response, which fits how v2 renders that text. The related tests pass. I couldn't check the actual effect on model behaviour, since there are no provider credentials. The public API is unchanged: only built-in prompt text changes. No further concerns.