Self-referential skill invocations do not render per harness

[williamthorsen]

Problem

A skill or subagent that names its own invocation in prose — e.g. capture-feedback's "Invoke /capture-feedback whenever the user gives feedback" — is written as a literal /slug. That renders correctly only on Claude, whose skill sigil is /; on a harness with a different sigil (Rovo's !) it stays /slug, which is wrong. These references can't adopt the {skill:} invocation token today because a self-targeting token makes the resolver treat the artifact as its own dependency and trips the cycle check.

The same literal-/slug spelling was also used for references naming a different skill. Those need reclassifying, not the same fix.

Context

#879 tokenized genuine cross-artifact invocations and left every literal-/slug reference alone, because a self-token forms a resolver cycle. The refined treatment splits the untokenized references three ways: a self-reference renders but must not depend on itself; a reference naming a runnable command in another artifact should ship that command (a genuine edge, not a render-only reference); and a reference pointing backward up an existing dependency — a depended-upon skill naming its depender — is not a dependency at all and would cycle if tokenized.

Proposed solution

Let an artifact render its own invocation per harness without contributing a dependency edge: the resolver drops a direct self-edge arising from a body invocation token, instead of throwing the cycle error. Genuine transitive cycles (A→B→A) and frontmatter-declared self-dependencies still error — only the direct body-token self-edge is dropped. (Drop the self-edge; do not add a separate render-only token.)

Apply the correct treatment to every literal-/slug reference across the library:

• Self-reference → {skill:self} / {subagent:self} body token: renders per harness, self-edge dropped, no dependency.
• Non-invoking cross-reference → {skill:other} token: a genuine dependency edge, so the named command ships with the referring artifact.
• Backward reference — a depended-upon skill naming its depender (capture-event → capture-feedback, create-devlog → wrap-up) → reword to plain prose; not a dependency, must not render as a command.

Acceptance criteria

Must have

• A body invocation token targeting its own artifact renders per harness and contributes no dependency edge.
• A genuine dependency cycle (A→B→A) and a frontmatter-declared self-dependency still raise the cycle error.
• Every literal-/slug self-reference in the library is converted to a self-token and renders per harness.
• Every literal-/slug cross-reference either becomes a dependency edge or, where it points backward, is reworded to non-command prose.
• New and modified behavior is covered by tests.