skills: one command per operation

[mnkprs]

Why

The agent interface was two skills: /fairtask and a /fairtask-eval router that took free text and picked an operation from a table. Distinct commands are easier to discover, invoke and document than "/fairtask-eval …".

What changed

• Five new skills, one operation each: /fairtask-baseline (screen a task with the one-prompt baseline for comparison — its verdict is explicitly labelled as not machine-verified), /fairtask-report (the two results tables, with a workspaces preflight), /fairtask-score, /fairtask-cases (the 30-case set, or one instance laid out with its decisive lines quoted), /fairtask-trajectory.
• /fairtask-eval becomes the umbrella for the supporting checks (evidence audit, zero-LLM code check, provenance) and routes everything else to the dedicated commands.
• README §2b: a command table after the install row; the eval examples now show the dedicated commands.
• Manifest validator covers all seven skills; CI plugin check expects Skills (7); plugin.json description lists the command set.