feat(cli): add rozenite skills command and flatten agent skill docs

[V3RON]

Description

Restructures Rozenite's agent skills so all real content is CLI-bundled and served by a new rozenite skills command, and the standalone installable skill becomes a thin router.

• Added rozenite skills list and rozenite skills show <id> to the CLI, backed by a small registry (packages/cli/src/skills/registry.ts) that reads bundled docs from packages/cli/docs/ and exposes lookup by doc id and by domain token.
• Flattened all skill content into a namespace of short, self-contained docs bundled in packages/cli/docs/: core (shared ground truths), cli (shell-driven workflow), sdk (programmatic workflow), sdk-patterns (copy-pastable code patterns), and one doc per agent-enabled domain (storage, mmkv, console, network, network-activity, react, react-navigation, redux-devtools, performance, memory, file-system, controls, tanstack-query).
• Replaced the rozenite-agent and rozenite-agent-sdk installable skills with a single rozenite skill (packages/cli/skills/rozenite/SKILL.md) that only tells the agent to run npx rozenite skills list/show <id> and start with core — it never hardcodes the doc inventory, so it can't go stale.
• Updated website/src/docs/agent/{skills,overview,sdk}.mdx to document the new skill and command.
• Added a changeset for the rozenite package.

No MCP server or rozenite mcp command is included — that is a follow-up. No version-check behavior was touched.

Related Issue

N/A — internal restructuring of existing agent-skill tooling.

Context

The two previous skills (rozenite-agent, rozenite-agent-sdk) embedded all their content directly inside the installable skill package, including domain references and code-pattern references between files. That made the skill large, coupled its content to the skills CLI's installation format, and required cross-file references between skill documents. Moving the content into the CLI package and exposing it via rozenite skills means the docs can be read on demand (npx rozenite skills show <id>), each doc is self-contained (no cross-references), and the installed skill itself only needs to change if the discovery mechanism changes, not when doc content changes.

The registry resolves its docs directory the same way TEMPLATE_DIR resolves template/ for init (relative to the built CLI entry point via import.meta.url), so it works from source and from the published package.

Testing

Automated:

• pnpm --filter rozenite run typecheck
• pnpm --filter rozenite run lint
• pnpm --filter rozenite run test (94 tests passed, including new skills-registry.test.ts covering: all expected doc ids are listed, every domain doc has its domain field set (and non-domain docs don't), domain-token lookup works, unknown id/domain lookups return undefined, and no doc body contains a cross-reference to another doc's filename or a references/ path)
• pnpm lint:all, pnpm format:all, pnpm typecheck:all (all packages pass)
• pnpm release:plan (confirms the changeset is picked up)

Manual:

• node packages/cli/bin.js skills list — prints all 17 docs with id/description (and domain where applicable)
• node packages/cli/bin.js skills show core — prints the real doc body
• node packages/cli/bin.js skills show does-not-exist — exits non-zero with a message listing valid ids

[V3RON]

Addressed review feedback:

1. Removed the misleading "memoized per directory" claim from loadAllDocs's doc comment in packages/cli/src/skills/registry.ts — it never memoized, and now the comment just says it reads from disk.
2. rozenite skills list now writes compact JSON by default with an opt-in --pretty flag, matching the agent command convention.
3. core.md now says npx rozenite skills show cli / npx rozenite skills show sdk instead of bare cli/sdk ids, so it's actionable without ambient knowledge of the fetch mechanism.
4. Extended the cross-reference test in packages/cli/src/__tests__/skills-registry.test.ts to reject doc ids adjacent to docs/, domains/, or references/ markers (in addition to <id>.md). Also added a dedicated unit test for the matching helper itself, since a literal "any slash adjacent to a known id" rule produces false positives against core's legitimate evil/mmkv / avasapp/ably domain-token-provenance examples and sdk's ./sdk npm subpath-export mentions — both are prose about naming conventions, not doc-location references.
5. website/src/docs/agent/skills.mdx now states plainly that the guidance lives in the rozenite CLI package and is versioned with it (so it can't drift from the installed Rozenite version), that the installed skill carries no workflow content of its own, and that npx rozenite skills list/show <id> work identically without installing the skill at all.

Re-ran pnpm --filter rozenite run typecheck/lint/test (95 tests pass), pnpm lint:all/typecheck:all, and the manual checks: node packages/cli/bin.js skills list (compact), skills list --pretty, skills show core, and skills show does-not-exist (non-zero exit, valid-id list).

[V3RON]

Smoke tested ✅

Ran the restructured skills + CLI end to end against a live app (iPhone 17 Pro simulator, playground app) via a Haiku sub-agent, using only the bundled docs (no prior Rozenite knowledge):

• Skill discovery: npx rozenite skills list, skills show core, skills show cli — all returned correct, readable content.
• Live round trip: agent read skills show controls to learn the domain's tool names, created a session (agent session create), toggled the mockLatency control in the simulator via agent-device, then called agent controls call --tool get-item ... and confirmed the CLI read back value: true — matching the on-device state.

Also verified separately: build + unit tests pass, npm pack --dry-run includes docs/ and skills/rozenite/SKILL.md, and found & staged a fix for a dangling .agents/skills/rozenite-agent symlink left over from the old skill layout (staged locally, not yet pushed: points at packages/cli/skills/rozenite).

Confirms the relocated skill router + CLI-bundled docs are self-sufficient for a low-capability model to discover and drive correctly.