feat: invoke a trusted skill for one console task

[AetherAI3]

Summary

• Add /skill <qualified-id> <task> to the idle raw console. Parse only the command and identifier; submit the complete remaining task as literal text through the existing resolver and RunSession. The selected skill is scoped to that turn, including when ordinary tasks were already queued.
• Refuse busy, queued, line-input, --no-skills, and server-executed cloud chat invocations before an unrestricted model call. Show the resolved ID, digest, context size, and host tool limits; /context next /skill ... previews a fresh selection without admitting it.
• Add a shared strict store reader for skill settings and trust, meeting [feature] Console: invoke a trusted skill for one task without restarting #327's Do not treat incompatible skill stores as empty and overwrite them #192 prerequisite: malformed/future stores refuse reads and mutations without changing their bytes. Discovery reads both stores once before skill admission.

Evidence

• npm run build passed.
• npm run docs:check passed; 6 generated outputs match.
• 123 focused tests passed across skill resolution and host policy, store corruption and trust, the raw composer and queue, command manifest and picker, and context inspection. A final 12-test focused run passed after the last inspector assertion.
• git diff --cached --check passed before commit.
• The raw TTY fixture proves literal / and ! task text reaches the local model with the skill, while a normal queued turn has no inherited skill. Other fixtures prove local model failure and disabled-skill refusal do not make selection sticky, and cloud/--no-skills refusals make zero model requests.

Closes #327
Refs #192