fix(tui-gateway): merge agent output on personality-marker history race

[pierrenode]

Summary

• config.set's "personality" key applies immediately regardless of session["running"], unlike "model" (which defers to next-turn-start while a turn is in flight — see the session.get("running") check in config.set's key == "model" branch).
• A personality change mid-turn (e.g. via the desktop/web settings UI while a response is streaming) appends a marker to session["history"] and bumps history_version — the exact same shape of mutation a model switch makes.
• The turn-completion mismatch guard in _run_prompt_submit (added by [Bug]: Model switch mid-session triggers history_version mismatch — all subsequent agent output discarded (empty assistant messages in DB) #76870, three days ago) only recognized model-switch markers as a safe, mergeable mutation. A personality-marker-only mismatch fell into the "genuine desync" branch instead: the agent's response for that turn was silently discarded from persisted history (only shown transiently with a "not saved to session history" warning).

Root cause

_apply_personality_to_session (tui_gateway/server.py) has no session.get("running") guard on the config.set RPC path — it calls straight through, appending {"role": "user", "content": marker} and incrementing history_version under session["history_lock"], whether or not a turn is currently in flight.

_run_prompt_submit's post-turn history_version check (introduced by #76870 for the analogous model-switch race) only ever compared against _is_model_switch_marker, so a personality marker landing mid-turn was treated identically to a genuine external desync (/undo, /compress, /retry) — the agent's output for that turn was dropped instead of merged.

Fix

Extend the merge-recovery check to also recognize personality-pivot markers as safe to merge, using the exact content-diff technique #76870/cc04825c5d already established and reviewed for model-switch markers:

• Added _is_personality_marker() (mirrors _is_model_switch_marker()), matching either the marker's stable "[System: ..." content-prefix (today's shape) or display_kind == "personality_switch" (the structured shape an unrelated open PR, fix(history): stop personality marker from swallowing the next real prompt #74350, would give newly-appended markers — checking both keeps this correct regardless of merge order between the two fixes).
• _run_prompt_submit's mismatch branch now strips both model-switch and personality markers before the content-diff comparison, and accepts either marker type as the sole safe mutation.

Test plan

• Added test_prompt_submit_merges_on_personality_marker, mirroring the existing test_prompt_submit_merges_on_model_switch_marker harness. Covers both cases (no prior personality marker, and a prior marker already in the turn-start history) — simulates the race by appending the marker from inside the stub agent's run_conversation(), exactly as a concurrent config.set RPC would mid-turn.
• Mutation-verified: reverted the tui_gateway/server.py change and confirmed the new test fails with the exact pre-fix symptom (agent output NOT written to session history, 0 assistant messages merged).
• tests/test_tui_gateway_server.py full suite: 517 passed, 1 pre-existing unrelated failure (test_write_json_serializes_concurrent_writes, a thread-scheduling-timing test — confirmed flaky/pre-existing by re-running it in isolation both with and without this change, unaffected either way).
• ruff check clean on both changed files.

Competitor check

• No open/closed PR targets this specific race (searched multiple keyword combinations against tui_gateway/server.py's config.set/_apply_personality_to_session/_run_prompt_submit merge-recovery logic).
• Two adjacent-looking open PRs were read in full and confirmed to be different bugs, not duplicates:
  • fix(history): stop personality marker from swallowing the next real prompt #74350 restructures _apply_personality_to_session to persist the marker durably with structured display_kind/display_metadata, fixing a different bug (the marker swallowing the next real prompt on resume/reload via alternation-repair). It never touches _run_prompt_submit's mismatch-merge block. This PR's _is_personality_marker() is written to stay correct whether fix(history): stop personality marker from swallowing the next real prompt #74350 lands before or after it.
  • fix(gateway): pivot personality on the next turn, not the one after #73708 adds an equivalent personality-pivot marker to the separate gateway/run.py / gateway/slash_commands.py path (a different gateway surface entirely, not tui_gateway/server.py).