readme: agent interface as its own section; one-row-per-command reference

[mnkprs]

Why

For an agentic-workflows submission, "use it from inside your agent" was a short subsection under the shell usage, and the command reference hid flags in a third column instead of showing complete commands.

What changed

• §2a "Use it from a shell": the previous shell usage, unchanged in substance.
• §2b "Use it from inside your agent", now a full section: the three install paths; /fairtask with its three inputs and what happens for each (including the fail-closed PR flow), a verbatim in-session verdict from examples/psf__requests-2317/, how the engine is located and installed at the pinned release, manual mode; /fairtask-eval with seven example invocations and what each runs and shows; what the skills may and may not do (read-only tools, workspace guard, cost, triage not decision).
• §2c "Command reference": one row per distinct invocation — every flag appears as a complete command with an explanation of what it does — grouped as screen / evaluate / run / data / checks.
• "Where to go" rows and cross-references updated to the new section ids.