fix(skills-evals): pin Claude 2.1.150, inline SKILL.md for copilot, rename copilot fixture - #47

2.1.156 (current @latest) emits a `system`-role message at
messages[1] of the wire payload. DeepSeek's Anthropic-compat
endpoint rejects it with `400 Failed to deserialize the JSON
body into the target type: messages[1].role: unknown variant
'system', expected 'user' or 'assistant'`.

Reproduces on both claude workflows against the post-#44 main:

  https://github.com/stackific/stax/actions/runs/26618513746
  https://github.com/stackific/stax/actions/runs/26618519248

2.1.150 is the most recent release I can verify doesn't carry
the regression (chosen by working back from 2.1.156; the gap
between 2.1.150 and 2.1.156 is six patch releases). Pinning
rather than @latest matches the policy the install-step comment
has always claimed.

Bump intentionally after verifying the smoke test still passes.

Signed-off-by: Tanzim Saqib <[email protected]>

Copilot CLI's `-p <prompt>` does NOT resolve leading slash commands the
way Claude Code does. Running `copilot -p "/scope build a single HTML
todo list app"` passes the prompt verbatim and the model treats it as
a literal request to build the app — skipping the scope planning step,
never creating any work-item plan files. Direct evidence from run
26618829987 (copilot user-scope, fix/copilot-workspace-fixture):

  [driver] initial_prompt: /scope build a single HTML and localStorage-
    based todo list app
  [driver] spawn: copilot -p /scope build a single HTML ...
  [driver] stdout #5: Done. Created `todo.html` — a single-file ...

The fixture-rename commit (b254fa8) made the collection-time error go
away and exposed this deeper, pre-existing pathology that has kept all
three copilot scenario tests (todo / work_item_extends /
reminders_supersedes_todo) failing since they were introduced in #10.
None have ever passed in CI; the prior fixture bug masked it as a
collection error rather than a scenario error.

Fix mirrors the cline driver's `compose_skill_prompt` strategy (which
itself dates back to opencode's pre-`--command` era): detect a
`/scope` or `/ship` prefix in the prompt, read the corresponding
SKILL.md off disk, prepend its body, append the same CI directive
every other inlining driver uses, and hand the composed text to
`copilot -p`. The model then sees the actual skill instructions and
follows them instead of treating the slash as ornamental.

Discovery paths probed (project then user) mirror copilot's
documented skill lookup list (constants.go: `.agents/skills/`,
`.claude/skills/`, `.github/skills/` at project; `~/.copilot/skills/`,
`~/.agents/skills/` at user) plus `~/.claude/skills/` because copilot
tests install via `stax init --agents claude` (the transitional shape
in conftest.AGENT_INIT_VALUE_FOR_KEY) which lands skills under
`.claude/skills/`.

Non-slash prompts (the smoke test's `Respond with the single word:
ok`) pass through unchanged so the no-skill code path keeps working;
`run.initial_prompt` retains the original slash form so logs and
transcripts read the way the test author wrote them.

CI_DIRECTIVE is copied verbatim from cline_driver — kept verbatim, not
imported, because each driver's docstring asserts cross-driver parity
is by-value and a `from cline_driver import CI_DIRECTIVE` would couple
the modules in a way that hides accidental drift in code review.

Signed-off-by: Tanzim Saqib <[email protected]>