Add Copilot as a SkillOpt-Sleep model backend (CopilotCliBackend) + research-engine MCP plugin

[Dongbumlee]

What this adds (and how it differs from existing Copilot support)

The repo already ships a Copilot MCP client plugin (plugins/copilot/mcp_server.py) that lets the Copilot CLI trigger a SkillOpt-Sleep cycle. But until now the cycle's actual LLM work could only run on the mock, claude, or codex backends — there was no CopilotCliBackend, so Copilot could orchestrate a sleep cycle but not run one on itself.

This PR adds Copilot as a model backend (backend="copilot"), closing that loop:

Layer	Before this PR	After
Copilot as orchestrator (MCP client plugin)	✅ exists	✅ unchanged
Copilot as model backend (runs attempt / judge / reflect)	❌ missing (mock|claude|codex only)	✅ backend="copilot"

It is also the only Sleep CLI backend that is explicitly Windows-safe (UTF-8 decoding + cross-platform tool shims), where claude/codex are Unix-oriented.

Summary

Adds first-class GitHub Copilot support to SkillOpt, in two independent pieces:

1. CopilotCliBackend for SkillOpt-Sleep — lets the self-evolution engine run its attempt / judge / reflect loop on the GitHub Copilot CLI, alongside the existing Claude and Codex CLI backends.
2. A SkillOpt research-engine MCP server plugin (plugins/copilot/skillopt/) — exposes skillopt_list_configs / skillopt_train / skillopt_eval to Copilot. This is a new, separate plugin from the existing SkillOpt-Sleep MCP server (plugins/copilot/mcp_server.py): it drives the research training/eval scripts rather than the Sleep cycle.

Why

SkillOpt-Sleep already supported claude and codex CLI backends, and a Copilot MCP client plugin already let Copilot trigger a cycle — but Copilot itself could not be the backend that runs the cycle. This extends the same CliBackend contract to the GitHub Copilot CLI so a Copilot user can drive validation-gated skill optimization end-to-end on Copilot, without switching tools.

What changed

Copilot backend (skillopt_sleep/backend.py)

• New resolve_copilot_path() + CopilotCliBackend, registered in get_backend() with aliases copilot, github_copilot, copilot_cli, gh_copilot. (Upstream backend.py had no Copilot backend; --backend choices were mock | claude | codex.)
• Invocation: copilot -p <prompt> --output-format json --stream off --no-color --log-level none --allow-all-tools -C <tempdir>, parsing assistant.message events from the JSONL stream. (-s/--silent returns empty stdout on Windows, so JSONL parsing is required.)
• Cross-platform attempt_with_tools: honest tool-call detection mirroring the Claude/Codex backends — writes per-tool executable shims into the work dir and detects real invocations from a calllog (not self-reported markers). Shims are cross-platform: a .cmd batch shim on Windows and a chmod'd bash shim on POSIX, with an OS-specific tool hint.
• Windows UTF-8 fix: force encoding="utf-8", errors="replace" on the subprocess — text=True decodes as cp1252 on Windows and crashes on Copilot's UTF-8 output (byte 0x9d).
• Startup optimization (~4.9×, 36s → 7.4s per call): runs each Copilot invocation against an isolated COPILOT_HOME (no user MCP servers spawned — avoids recursively launching SkillOpt's own MCP servers), plus --disable-builtin-mcps and --no-custom-instructions. Auth is unaffected (stored in the OS credential store).
• Escape hatches via env: SKILLOPT_SLEEP_COPILOT_HOME, SKILLOPT_SLEEP_COPILOT_MODEL, SKILLOPT_SLEEP_COPILOT_FULL_ENV=1.
• copilot added to --backend choices in __main__.py and experiments/run_experiment.py; the config comment and the existing Sleep MCP server's backend enum are updated to match.

Research-engine MCP plugin (plugins/copilot/skillopt/)

• New, stdlib-only mcp_server.py driving scripts/train.py / scripts/eval_only.py, plus mcp-config.example.json, an instructions snippet, and a README — same structure as the existing plugins/copilot/ SkillOpt-Sleep server, but for the research loops. (No research-engine MCP plugin existed upstream.)

Tests (tests/test_sleep_engine.py)

• TestCopilotBackend (7 tests, no real CLI required):
  • _parse_jsonl_response — multi-message concat, junk-line skipping, empty / non-assistant events
  • get_backend alias resolution
  • isolated-home / full-env / home-override behavior
  • attempt_with_tools honest detection via an OS-aware stub (runs on both Windows and Linux)

Validation

• Full suite: 104 passed, 2 skipped (the 2 skips are pre-existing and data-dependent).
• Lint: ruff check reports 0 new findings (pre-existing findings on main are unchanged).
• Real-data run via experiments/run_experiment.py with the copilot backend: researcher persona showed a 0.42 → 1.00 lift; the programmer persona's validation gate correctly rejected non-improving edits.
• Cross-platform attempt_with_tools verified live: real Copilot call on Windows (tool actually invoked, detected from the calllog) and the POSIX path on Linux/WSL.

Notes

• Two separate backend systems exist in the repo: the research skillopt/model/ system (async ModelBackend) and the SkillOpt-Sleep skillopt_sleep/backend.py system (sync CliBackend). This PR's Copilot backend follows the Sleep CliBackend pattern (alongside ClaudeCliBackend / CodexCliBackend); the research-engine MCP plugin simply shells out to the existing scripts/ entry points.