hermes - 💡(How to fix) Fix gateway: pre-output message assimilation with tool-dispatch gate (follow-up to #31341)

Background

#31341 landed B1.2 active-session text debounce — rapid follow-ups during a turn are now merged into the next turn instead of being clobbered or fragmenting. That solves the correctness problem.

Remaining gap

When the agent has started a turn internally but hasn't yet produced visible output or dispatched tools, a follow-up message still results in two turns: response-to-msg1, then response-to-msg2+msg3. With assimilation, this could collapse into one turn where the agent responds to all three messages together.

This is an optimization, not a correctness fix — the user-visible behavior is "2 turns instead of 1" rather than anything broken.

Why my previous attempt (#29943) was wrong

The timing-window approach didn't gate on side effects. Cancel/restart could discard a turn after a tool had already executed (file write, network call, terminal command). Shipping behind an off-by-default flag confirmed it wasn't ready.

Proposed approach: tool-dispatch gate

Replace the timing window with a state machine on the agent run context:

• Track committed: bool on the run context
• Set True on first tool dispatch (before execution) OR first visible output token
• Adapter checks is_turn_committed(session_key) before attempting assimilation
• Assimilation only allowed while committed == False
• No timing deadline as correctness gate — only as a staleness hint for stale revisions
• Ship default-on once correct (no `HERMES_PRE_OUTPUT_*` flag needed)

The key shift: "no side effects until commit" instead of "race to assimilate before side effects."

Question for maintainers

Is this optimization worth pursuing as a follow-up to #31341, or is the 2-turn case acceptable given the user value B1.2 already delivers?

If yes, I'd implement in a focused PR with the gate in `gateway/run.py` and a thin adapter-side check — no large state expansion in `BasePlatformAdapter` this time.