feat(skills): add a body-owned SKILL.md workflow layer

[brandwe]

Problem

Reusable workflows currently require edits to the always-loaded body prompt or repeated ad hoc instructions. That increases prompt size, requires a server restart, and gives sponsors no supported way to maintain small, reusable operating recipes.

Desired outcome

Add a body-owned skills layer that discovers trusted SKILL.md recipes and activates one explicitly for a single turn when a human invokes it.

Recommended priority: P1.

Scope

• Discover local and bundled skill roots.
• Parse bounded YAML frontmatter with safe loading and strict name validation.
• Support manual /skillname invocation plus /skill list|show|refresh|help.
• Make local skills override bundled skills by name.
• Reject symlinks, traversal, malformed metadata, reserved names, and oversized bodies.
• Audit every activation before emitting effective prompt context.
• Keep persona-sati observation behavior unchanged.

Non-goals

• A skills marketplace.
• Executing scripts or loading arbitrary sidecar files from a skill directory.
• Agent-initiated semantic routing in the MVP.
• Letting skills override the non-overridable body prompt.
• Adding tools or MCP servers dynamically.

Acceptance criteria

• Valid local and bundled skills load deterministically; local precedence is tested.
• Manual invocation forwards arguments and injects only the selected skill for one turn.
• Special commands return bounded, Teams-safe output.
• Unknown slash commands do not activate a skill.
• Symlink, traversal, unsafe YAML, malformed metadata, reserved-name, and size-limit cases fail safely.
• Activation is audit-first and fails closed if the audit write fails.
• The feature is optional and Entrabot boots normally with no skills directory.
• Documentation and tests match the plan; full pytest and Ruff gates pass.

Sources and related work

• Detailed skills plan
• Master feature plan
• XPIA boundary work is already shipped and remains authoritative for untrusted external content.