fix(mcp): wire run_powershell and MCP server for Windows/Scoop environment

- Add .mcp.json at project root (correct location for claude mcp add)
- Add mcp_env.toml: project-scoped PATH/env config for subprocess execution
- shell_runner.py: load mcp_env.toml, add stdin=DEVNULL to fix git hang
- mcp_server.py: call mcp_client.configure() at startup (fix ACCESS DENIED)
- conductor skill files: enforce run_powershell over Bash, tool use hierarchy
- CLAUDE.md: document Bash unreliability on Windows, run_powershell preference

.claude/commands/conductor-implement.md

@@ -24,11 +24,14 @@ Follow this EXACTLY per `conductor/workflow.md`:
 Edit `plan.md`: change `[ ]` → `[~]` for the current task.
 
 ### 2. Research Phase (High-Signal)
-Before touching code, use context-efficient tools:
-- `py_get_code_outline` or `py_get_skeleton` (via MCP tools) to map architecture
-- `get_git_diff` to understand recent changes
-- `Grep`/`Glob` to locate symbols
-- Only `Read` full files after identifying specific target ranges
+Before touching code, use context-efficient tools IN THIS ORDER:
+1. `py_get_code_outline` — FIRST call on any Python file. Maps functions/classes with line ranges.
+2. `py_get_skeleton` — signatures + docstrings only, no bodies
+3. `get_git_diff` — understand recent changes before modifying touched files
+4. `Grep`/`Glob` — cross-file symbol search
+5. `Read` (targeted, offset+limit only) — ONLY after outline identifies specific ranges
+
+**NEVER** call `Read` on a full Python file >50 lines without a prior `py_get_code_outline` call.
 
 ### 3. Write Failing Tests (Red Phase — TDD)
 **DELEGATE to Tier 3 Worker** — do NOT write tests yourself:
@@ -48,6 +51,7 @@ Run tests. Confirm they PASS. This is the Green phase.
 With passing tests as safety net, refactor if needed. Rerun tests.
 
 ### 6. Verify Coverage
+Use `run_powershell` MCP tool (not Bash — Bash is a mingw sandbox on Windows):
 ```powershell
 uv run pytest --cov=. --cov-report=term-missing {TEST_FILE}
 ```