fix: fix runtime detection priority, resolve binaries from APM runtimes dir, add llm support

[edenfunf]

Summary

Fixes apm run start selecting the wrong runtime or failing to execute one on Windows, and adds llm as a supported runtime for users who cannot use codex with GitHub Models.

Closes #605.

Background — codex v0.116+ and GitHub Models

codex removed wire_api = "chat" (Chat Completions) in v0.116 and now only supports wire_api = "responses" (OpenAI Responses API). GitHub Models exposes only the Chat Completions endpoint — the /responses path returns 404. This is a fundamental upstream incompatibility; there is no config workaround.

apm runtime setup codex always installs the latest release, which is now in the incompatible range. The setup scripts are updated to document this so users are not silently directed to a broken binary.

Changes

src/apm_cli/core/script_runner.py

Runtime detection (_detect_installed_runtime)

• Check ~/.apm/runtimes/ before PATH. A system-level copilot shim (e.g. from gh extension install github/gh-copilot) was winning over an APM-managed binary because shutil.which() searched PATH first.
• Exclude APM-managed codex from auto-detection. Any codex installed via apm runtime setup will be v0.116+ and will not work with GitHub Models. PATH codex is still tried as a last resort (covers older or differently-configured binaries).
• Prefer llm over other PATH runtimes. llm uses Chat Completions and is fully compatible with GitHub Models via the GITHUB_TOKEN env var.

Executable resolution (_execute_runtime_command)

• On Windows, resolve the binary path from ~/.apm/runtimes/ before falling back to shutil.which(). Without this, the correct runtime name was detected but the executable could not be found if ~/.apm/runtimes was not in the session PATH.

llm support (_generate_runtime_command)

• Add llm -m github/gpt-4o <prompt_file> as the generated command for the llm runtime.

Additional cleanup

• Remove anchored ^ regex in _get_runtime_name and _transform_command; runtime keywords can appear mid-string when an absolute path is prepended on Windows.
• Remove the blanket symlink rejection in _discover_prompt_file and _resolve_prompt_file; it was too broad and blocked legitimate use cases (e.g. symlinked prompt packages).
• Improve the "prompt not found" error message with actionable steps.

scripts/runtime/setup-codex.ps1 / setup-codex.sh

• Document the codex/GitHub Models incompatibility in a comment so operators know why GitHub Models is not listed as a working provider for current codex builds.

Test Plan

• uv run pytest tests/unit/test_script_runner.py -x -v — all tests pass
• apm runtime setup codex followed by apm run start — selects llm (or copilot) instead of the broken codex binary
• With only llm available in PATH and no APM-managed runtimes: apm run start — llm -m github/gpt-4o is executed
• With a system PATH copilot stub and a working APM-managed copilot binary: apm run start — APM-managed binary wins
• Absolute path to runtime binary works on Windows when ~/.apm/runtimes is not in PATH