fix(llm): honor SIMARD_LLM_PROVIDER / AMPLIHACK_LLM_PROVIDER env override

[rysweet]

Why

Simard's OODA daemon was scoring 0.00% on every L1-L12 evaluation while reporting ✓ completed — a textbook hollow-success bug. Root cause: amplihack's LLM router defaults to the bundled Claude Code CLI, which is "Not logged in" by default. The daemon's SIMARD_LLM_PROVIDER=copilot env var was being ignored.

Two intertwined bugs.

Bug 1: _detect_launcher is purely file-based

amplihack.llm.client._detect_launcher reads <project_root>/.claude/runtime/launcher_context.json and returns "claude" when the file is absent, stale, or malformed. Embedded callers that don't go through amplihack copilot — Simard's Rust daemon imports amplihack.eval directly via Python subprocess — never have a launcher context written, so detection unconditionally falls through to "claude". The SIMARD_LLM_PROVIDER env var is never read.

Bug 2: copilot SDK probe imports a non-existent module

from copilot.types import MessageOptions, SessionConfig  # ModuleNotFoundError

copilot.types doesn't exist in copilot >= 0.1.0. That ImportError silently set _COPILOT_SDK_OK = False, making the entire copilot path dead code. _query_copilot was also calling a stale API (client.start(), session.send_and_wait(MessageOptions(prompt=...))) that no longer exists.

Fix

Selector (_detect_launcher):

1. Explicit env override (AMPLIHACK_LLM_PROVIDER, then SIMARD_LLM_PROVIDER) — wins unconditionally
2. File-based detection via LauncherDetector
3. Default "claude"

Recognized provider values: claude/anthropic/claude-code and copilot/github-copilot/gh-copilot/rustyclawd. Unknown values fall through to file detection.

Dispatch (completion): when an env override is present, refuse to silently fall back across providers. If the user pinned copilot but the SDK is unavailable, log a warning and return "" rather than masking the misconfiguration by routing to Claude.

Copilot SDK (_query_copilot): rewritten against the real copilot 0.1.0 API:

async with CopilotClient() as client:
    session = await client.create_session(
        on_permission_request=PermissionHandler.approve_all,
        working_directory=str(project_root),
    )
    event = await session.send_and_wait(prompt, timeout=float(QUERY_TIMEOUT))

Verification

End-to-end smoke against the installed copilot SDK:

$ SIMARD_LLM_PROVIDER=copilot python -c "
import asyncio; from amplihack.llm import completion
print(asyncio.run(completion(messages=[{'role':'user','content':'Reply with just the word: OK'}])))
"
OK

Tests

17 new tests in tests/llm/test_provider_env_override.py:

• Value normalization (case, whitespace, aliases)
• AMPLIHACK_LLM_PROVIDER priority over SIMARD_LLM_PROVIDER
• Env override beats existing launcher_context.json on disk
• Unknown values fall through to file detection
• No-silent-fallback contract: pinning copilot must NOT route to Claude if copilot SDK is unavailable, and vice versa

All 19 LLM-related tests pass (17 new + 2 existing in tests/eval/test_llm_grader.py).

Impact

Once merged and deployed via uv tool upgrade amplihack, Simard's OODA cycles should produce real, non-zero L1-L12 scores instead of the current 0.00% hollow-success regime. Further, any caller that sets SIMARD_LLM_PROVIDER=copilot or AMPLIHACK_LLM_PROVIDER=copilot in the environment now gets the behavior that name advertises, with no .claude/runtime/launcher_context.json ritual required.