MCP `steer`: say a sentence to a running agent run, with the same rules as the API #5213

Done: MCP steer { repo, run, text } steers a running agent run, through the same action the API uses (steerRepoRun, POST …/runs/:id/steer), so its rules, refusals and access are exactly the route's — no second implementation.

• [ ] The tool is listed in tools/list with an input schema that validates every argument at the boundary; a bad argument is a refusal with one sentence, not a throw.
• [ ] The answer is JSON with a says sentence, like the tools beside it.
• [ ] A row in docs/MCP.md's tool table.
• [ ] Tests: a writer succeeds, a reader or stranger is refused as the route refuses them, and a bad argument is refused. An empty text and a run that is not an agent's are refused with the route's words.

Files: src/http/mcp.ts, docs/MCP.md, one new test file. Risk: mcp.ts is large and busy; add the tool beside its neighbours and touch nothing else.

Fresh: replaces #4973. Both tries failed on tests it never touched (the burst-of-errors board, #5107; verify-email's agent test). This waits for those fixes.

duplicate of #4973