feat: clarify project manager continuations and skills

[numnx]

What changed

• Render agent-scheduled wakeups as a dedicated Project Manager continuation card in both Threads and 3D Chat.
• Show the next-step instruction, scheduled time when available, and queued, running, completed, or failed delivery state without presenting the event as a user-authored message.
• Inject the complete linked persistent-skill descriptor inventory so the Project Manager can accurately answer which skills are available before falling back to retrieval.
• Preserve versioned, read-only skill snapshots, revision provenance, search_skills retrieval, and manage_skills mutation routing from the current dev architecture.
• Remove obsolete generic subagent boilerplate from the built-in worker prompt.
• Update canonical and published documentation for both behaviors.

Why

Agent-owned scheduler wakeups were stored in the same message direction as dashboard prompts, so the UI could make an automated continuation look like something the user sent. Persistent skill attachments were also discoverable through search but were not summarized as an authoritative inventory, which made broad skill-availability questions easy to answer incorrectly.

User and developer impact

Scheduled follow-ups are now visually auditable and clearly attributed to the Project Manager. Project managers receive linked skill names, descriptions, and versions in runtime context while retaining vector-backed retrieval for detailed guidance. The branch is synchronized with current dev, including containerized Git storage and optimized memory retrieval.

Validation

• pnpm run lint
• pnpm exec vitest run tests/backend/services/skill-service.test.ts tests/backend/infrastructure/providers/cli/mockup-cli-provider.test.ts tests/dashboard/v2/chat-message-bubbles.test.tsx — 74 passed
• pnpm run test:backend — 5,077 passed, 7 skipped
• pnpm run test:dashboard — 1,991 passed, 2 skipped
• pnpm run sync:docs-web
• pnpm run check:docs-web
• pnpm run build

Risk and rollback

Risk is limited to chat presentation and additional descriptor text in skill-enabled agent prompts. Revert the three feature commits to restore prior rendering, prompt inventory behavior, and worker guidance independently.