CoderMCPServer Kind E2E follow-ups #390

Follow-ups from the final assessment of #389 (CoderMCPServer Kind E2E phases, M10). None of them blocked the merge.

1. Add offline failure rows for three checks that no test makes fail yet: "unlisted object wrote to Coder", "refused object wrote to Coder", and "no ControlPlaneGone log for the MCP server". The fake controller always prints that log.
2. A local rerun after a mid-run failure that keeps the namespace gets HTTP 409 on e2e-mcp-llm and e2e-mcp-taken. The older AI phases have the same problem with e2e-taken. CI always starts from a fresh cluster.
3. The CI failure dump in ci.yaml does not include the CoderMCPServer objects or the MCP stub pod logs.
4. The P1 leak check does not read the Coder pod log.

Refs #155.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: medium