feat(orchestrator): real ClaudeAgentHarness via the Claude Agent SDK — P1b-iv (7/7) - #1123

...(P1b-iv)
Makes the inner-execution seam real: ClaudeAgentHarness drives Claude Code for
one turn via the Claude Agent SDK, wired to teammcp. Env-var NAMES arrive on the
wire (McpServer.env_passthrough) and are resolved to values HERE from the
orchestrator's own environment, never on the wire — same trust model as
internal/provider/claude.go.
The intellectual content is a PURE outcome classifier: classify_outcome keys on
the agent's real team_task actions (grounded in the broker's vocabulary —
complete/done -> completed, submit_for_review -> review, block -> blocked,
planning-without-a-terminal-action -> plan_ready, else continue). Claude Code
namespaces MCP tools as mcp__<server>__team_task, so the classifier matches the
tool-name suffix. This keeps the policy fully unit-tested without a model key or
the SDK installed; only the thin SDK glue (build options, run query, collect
transcript) is lazy + pragma-no-cover.
build_harness degrades to FakeHarness when the SDK is absent, so the service
stays runnable key-free and existing tests are unchanged. This also closes the
P1b-iii FakeHarness gap: the real harness returns a real gate state instead of
re-projecting running.
Renumbered from 'P2' to P1b-iv to avoid colliding with the migration plan's P2
(the CEO decompose node); this work completes the single-task §5b seam.
- classify_outcome + TurnTranscript/ToolCall (pure, 11 cases incl. priority +
 mcp-prefix matching + non-team_task action ignored).
- _mcp_servers_config resolves env names from a provided env (value present,
 unset -> , ungranted vars absent).
- run_turn fails loud with install guidance when the SDK is absent (CI path).
- _prompt_for / _permission_mode_for (planning -> read-only 'plan').
- 43 pytest green; live smoke still SMOKE OK (degrades to FakeHarness).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>