feat(tui): loop-swap adapter — race genuinely different reasoning loops

[0bserver07]

Real backend heterogeneity — swap the reasoning loop

Phase 3.3 shipped per-lane preset + posture; this closes the gap I flagged as
deferred. A lane can now run a genuinely different reasoning loop — and the
multiplexer renders it like any other lane:

chimera code --tui --models glm-5.2,glm-5.2:coding_agent:plan-execute
#                           default ReAct    PlanAndExecute (a different loop)

Loops: plan-execute, reflexion, tot (+ the postures plan/tdd from 3.3).

The adapter

Only AgentLoop emits the LoopEvents the TUI renders; the strategy loops in
chimera/core/loops/ instead expose iter_steps() -> StepResult.
chimera/assembly/loop_adapter.py bridges them:

• runs the loop's synchronous generator in a worker thread, so its blocking
  provider calls don't stall the event loop and lanes stay concurrent;
• translates each StepResult (assistant text / tool calls / results) into
  LoopEvents, ending with exactly one terminal result (like AgentLoop);
• cooperative cancel at step boundaries via the agent's abort signal;
• wraps the provider so the loops' non-streaming complete() uses a bounded
  max_tokens — glm-5.2's large default otherwise trips the SDK's "streaming
  required" guard (which AgentLoop sidesteps via async_stream).

CodingAgent.run branches to the adapter for a real loop; postures and the
default stay on AgentLoop.

Verification

• tests/assembly/test_loop_adapter.py (7): translation, terminal result,
  cancel, error path, bounded provider. Suite 79 passed.
• Ruff clean; mypy chimera/ green (CI-parity, 674 files).
• Live on GLM-5.2: a ReAct lane and a plan-execute lane race one task; the
  swapped loop executes real tools, writes its file in its isolated worktree,
  and streams tool_use/tool_result/assistant into its pane.

🤖 Generated with Claude Code