feat(runtime): relay live executor output through streamAgentTurn

[drewstone]

Reproduction

Run any public executor through streamAgentTurn in @tangle-network/agent-runtime@0.131.0.
The executor's execute() returns an async iterable and emits usage while the underlying CLI Bridge or environment provider emits text and tool activity.

streamAgentTurn drains that iterable without yielding content:

if (isAsyncIterable(run)) {
  for await (const _usage of run) throwIfAborted(signal)
  result = executor.resultArtifact()
}

It then yields only llm_call, result-level tool calls, artifact, and final after execution settles.

Actual behavior

• CLI Bridge receives SSE text and tool events internally, but consumers see no text_delta, reasoning_delta, tool_result, or live tool activity.
• Provider executors collect AgentEnvironmentEvent[] internally, but consumers cannot observe them during the run.
• A terminal application cannot show the harness output while it is running.

Expected behavior

streamAgentTurn should project executor progress into RuntimeStreamEvent incrementally.
The shared executor progress type needs a lossless Runtime-owned event form for text, reasoning, tool calls/results, interactions, and provider events.
Usage events must continue through the existing accounting path.

This is a regression from the former chat and box paths, which emitted live normalized events.
It blocks Braid from rendering current CLI Bridge and Tangle sandbox activity without duplicating Runtime's parsers.

[drewstone]

Concrete live reproduction from Braid on the latest published packages:

• @tangle-network/agent-runtime@0.131.5
• @tangle-network/agent-provider-cli-bridge@0.4.3
• Pi 0.84.1 through CLI Bridge
• One real coding turn: the CLI Bridge trace recorded 7 Pi tool calls, while streamAgentTurn(..., { preserveToolParts: true }) emitted 0 tool_call and 0 tool_result events.

There is also a terminal shape mismatch in the current bridge path. streamBridgeSession appends step.toolName to out.toolCalls, producing string[]. executorResultToolCalls accepts only object entries with a string name, so it drops every bridge tool call even after the run finishes.

Fixing that shape would restore terminal call names, but this issue still needs the larger live projection and tool-result work described above. Acceptance for Braid: the same live Pi turn exposes ordered call and result events before final, and replay after completion retains them without reading CLI Bridge trace files.