#15518: tui: recover local no-output runs via history fallback

#### Summary Fixes a TUI reliability bug where local runs can finalize as `(no output)` even though assistant text is persisted shortly after. Code word: lobster-biscuit #### Repro Steps Related reports: - https://github.com/openclaw/openclaw/issues/4499 - https://github.com/openclaw/openclaw/issues/15432 1. Start `openclaw tui` with a provider/model that intermittently emits empty `chat.final` payload text. 2. Trigger a run that finishes with `(no output)` in TUI. 3. Reopen/reconnect TUI and observe assistant text appears from history. #### Root Cause `chat.final` can occasionally arrive without renderable assistant text while transcript persistence catches up moments later. TUI currently finalizes immediately and never retries history for local runs, so users see `(no output)` until reconnect. #### Behavior Changes - For **local runs only**, if final text resolves to `(no output)` and stopReason is not `error`, schedule two short history refresh retries (200ms, 1000ms). - Keep existing behavior unchanged for non-local runs and for runs with valid final output. - Centralize no-output placeholder handling in formatter helpers. #### Codebase and GitHub Search - Code paths searched: - `src/tui/tui-event-handlers.ts` - `src/tui/tui-stream-assembler.ts` - `src/tui/tui-formatters.ts` - Issue search checked: - #4499 (`MiniMax responses show '(no output)' until TUI restart`) - #15432 (`[Bug]: no output`) #### Tests - `pnpm vitest src/tui/tui-event-handlers.test.ts src/tui/tui-formatters.test.ts --run` - `pnpm check` #### Manual Testing ### Prerequisites - A TUI session connected to gateway. ### Steps 1. Run a local TUI conversation where a final event may be empty. 2. Confirm TUI schedules short history fallback and self-recovers without restart. 3. Confirm normal final outputs still render once with no extra history churn. #### Evidence - Added targeted test: local final `(no output)` triggers 2 history retries in `src/tui/tui-event-handlers.test.ts`. **Sign-Off** - Models used: GPT-5 Codex - Submitter effort (self-reported): High - Agent notes (optional, cite evidence): Local targeted tests + lint/type checks passed.