feat: expose final_message step output for defense-in-depth review workflows

[stanley1208]

Closes #1317.

Adds a final_message step output containing the agent's final text response, pulled from the last message in the SDK stream.

The use case the issue describes is read-only review workflows: run the AI job with contents: read only, then post the response from a separate job that has pull-requests: write. Right now you can't really do that. The only way to get the response text is to parse execution_file, which has no documented format. The openai/codex-action repo solves it with a final-message output, which is what the issue references.

Usage

- uses: anthropics/claude-code-action@vX
  id: review
  with:
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

- name: Post review
  env:
    REVIEW: ${{ steps.review.outputs.final_message }}
  run: gh pr review --comment --body "$REVIEW"

A note on naming

The issue uses final-message (kebab-case) because that's what codex-action uses. I went with final_message (snake_case) to match the other outputs in this repo (execution_file, structured_output, session_id). Easy to swap if you'd rather match codex-action.

How it works

A small helper called extractFinalAssistantMessage in base-action/src/run-claude-sdk.ts walks the message stream backwards, finds the last assistant message, and joins its text blocks. If the last action was a tool call (no text), the output is undefined. The intent is that "final message" should mean what the agent actually said.

Wired through both layers (the standalone base-action and the parent action). Added 10 unit tests for the edge cases: empty stream, multiple text blocks, text + tool_use mixed, tool-only last message, malformed content.

Checks

Ran bun test (10 new pass, no regressions), bun run typecheck, and bun run format:check.