v0.19.0 — a live, instrumented terminal UI

chimera tui goes from a single-pane chat to a streaming, instrumented workspace — reusing the shared ChatSession core, so nothing else changes.

Added

• Live token streaming + Markdown. The model's text streams into the log as it's generated (single-model path); finished replies render as Markdown with syntax-highlighted code. New LLMGateway.stream_complete() reassembles content + tool-call deltas + usage into a normal CompletionResult (the existing stream() primitive finally has a consumer).
• Activity panel — shows, from real signals, what the agent did this turn: the tools it called, the token count + cost, and how many memory facts were recalled.
• Ergonomics — a command palette (Ctrl+P), slash-command autocomplete, more key bindings (Ctrl+R reset, Ctrl+L clear, PgUp/PgDn scroll), a --stream/--no-stream flag, and a graceful fallback to chimera chat if Textual is missing.
• Agent.run now accepts on_token/on_tool callbacks and aggregates per-run token usage + cost into AgentResult; ChatSession.send_verbose() returns a TurnReport. All additive — the blocking complete()/send() paths are unchanged.

Honesty notes

• No token stream under --fuse (a panel→judge→synthesizer turn) — the panel shows a "synthesizing" status, not a fake cursor.
• Cost reads "unavailable" when the model's list price is unknown — never guessed.
• No verify/revert indicator in chat — verify-or-revert runs in solve/project, not in a chat turn, so promising it here would be dishonest.

Gate: 1433 passed / 3 skipped, mypy + ruff green. Live-smoked streaming (tokens + usage) against the OpenRouter free model.

Try it: `pip install --upgrade 'chimera-agent[full]'` then `chimera tui`