fix: preserve UTF-8 output for attached coding agents

[dgokeeffe]

What does this PR do?

• Preserves multibyte UTF-8 characters split across PTY reads.
• Decodes OpenCode SSE bytes explicitly as UTF-8.
• Enables app-SP OAuth for OpenCode and skips the PAT prompt in SP mode.
• Prefers rotated [DEFAULT] PATs over stale startup tokens.
• Prefers the loopback SP token broker when present, while retaining the existing fallback paths.
• Isolates MLflow autolog setup in unit tests.

Test evidence

• Combined stack make test: 459 passed, 1 skipped.
• Broker resolver regression coverage is included.
• Deployed to coding-agents-02 in the Lakemeter workspace as part of fix(host): close Omnigent final-review gaps #100 verification.
• Browser-driven PTY round-tripped café Ελληνικά 日本語 😀 without replacement characters or Latin-1 mojibake.
• Earlier Pi and OpenCode stress runs each produced 200 clean mixed-Unicode lines.

Pi conversation: conv_b7e103fe3d1b4b71b471d7c8e68da331

OpenCode conversation: conv_0ea7d5c71538423f9df1214d7191a69e