v0.19.1 — streaming-TUI hardening + memory-layer label

A tenth adversarial review, aimed at the brand-new v0.19.0 streaming surface — four gaps, no CRITICAL/HIGH (the tricky parts — usage accounting, thread marshalling, fusion honesty, markdown escaping — reviewed clean). All fixed and regression-tested.

Fixed

• TUI: a second Enter can't start a concurrent turn. The input is disabled while a turn is in flight (a thread worker can't be preempted, so a second submit would run a concurrent send_verbose on the same non-thread-safe ChatSession), and re-opens when the turn finishes.
• stream_complete degrades instead of erroring on strict providers. stream_options is now sent with drop_params=True, so a native anthropic/gemini model (or a strict api_base) that rejects it drops the param and streams anyway rather than failing every TUI turn.
• Streamed tool calls without a provider index no longer split in two — fragments merge into the open call instead of losing the arguments.
• Cost honesty: the panel flags "(excl. cache)" next to the turn cost when cache tokens are present.

Added

• The activity panel now names the memory layer that contributed — semantic / fts / keyword (or keyword+graph), shown alongside the fact count. Absent, not guessed, when unknown.

Gate: 1439 passed / 3 skipped, mypy + ruff green.

Upgrade: `pip install --upgrade 'chimera-agent[full]'`