ci(skills): lint, type-check and test the Python under skills/

[potiuk]

Summary

• Nothing checked the Python in skills/. The workspace checks iterate over [tool.uv.workspace] members, every member lived under tools/, so ruff, mypy, and pytest never saw skills/** — eight helper scripts and guards, plus the nine tests fix(pr-management-triage): centralize terminal PR links #1049 had just added, none of them run.
• Declares skills/ as one workspace member whose pyproject.toml is a config carrier, not a package (package = false). The existing machinery does the rest: run-workspace-check.sh discovers applicable checks from the sections present, and tests.yml derives its matrix from the members list, so this adds a pytest (skills) job with no workflow edit.
• Turning the checks on immediately found seven real things, all fixed here.

Closes #1053.

Type of change

• Skill change (.claude/skills/<name>/) — eval fixtures updated below
• Tool / bridge contract (tools/<system>/*.md)
• Python package (tools/*/ with pyproject.toml)
• Groovy reference impl
• Cross-cutting (RFC, AGENTS.md, sandbox, privacy-LLM)
• Documentation (docs/, README.md, CONTRIBUTING.md)
• Project template (projects/_template/)
• CI / dev loop (prek, workflows, validators)

Test plan

• prek run --all-files passes (23 hooks)
• ruff check, mypy, and pytest all run and pass through the workspace runner for the new member — uv run --project . python -m … from skills/
• Verified the coverage is real, not nominal: breaking the OSC 8 emission in pr_link.py now fails three tests under the workspace runner. Before this change the same break produced no signal at all — that is the whole point of the PR, so it seemed worth proving rather than asserting.
• check-workspace-members passes — it already scans top-level directories (widened in fix(ai-tutors): stop nested code fences inverting the fence state #1011), so skills is discoverable and not reported as drift.

RFC-AI-0004 compliance

• HITL — any new mutation is gated on explicit user confirmation
• Sandbox — no new unrestricted host access; network reach declared in the adapter
• Vendor neutrality — placeholders used in all skill / tool prose
• Conversational + correctable — agentic-override path documented if behaviour is adopter-tunable
• Write-access discipline — no autonomous outbound messages; drafts only, sent on confirmation
• Privacy LLM — private content does not reach a non-approved LLM; redactor invoked where needed

(None apply — CI wiring plus type/lint fixes.)

Linked issues

Closes #1053. Follow-up to #1049, which added the first tests under skills/, and to #1011, which fixed the same class of gap for ai-tutors/.

Notes for reviewers (optional)

Why one member and not per-skill packaging. Skills are symlinked into adopter repos one directory at a time (.claude/skills/magpie-<name> → skills/<name>), so a pyproject.toml inside each skill would leak build metadata into every adopter. skills/pyproject.toml is never symlinked — the relays point at skills/<name>, never at skills/ itself. I checked that before choosing this shape.

What the checks found. None of it changes behaviour; these are the annotations and renames the checks require, which is rather the point — they were invisible until now.

• scan_ci_runners.py — zip() without strict=; two same-named futures / future bindings of different types in one function, which is why the second pair needed renaming before mypy could annotate them; an unannotated rows; and step.get("with") called twice, so the isinstance narrowing never stuck.
• collect_status.py — l as a variable name, and a mixed-value dict literal inferred as dict[str, object], which made a later indexed assignment invalid.
• test_pr_link.py — a stale # noqa: E402 suppressing nothing.

One deliberate omission: ruff format is skipped for this member via [tool.magpie.checks] skip = ["ruff-format"]. The formatter wants to reflow ~280 lines across six pre-existing scripts — long set literals and regex constants — none of which this change touches. Folding that in would bury the substance behind a mechanical reformat nobody can review line by line. Lint, types and tests all run; formatting can land as its own commit and the skip be deleted then. Happy to do that immediately after if you would rather have it in one go.