feat(plugin): add databricks-experimental as a skills-only plugin

[simonfaltum]

Stacked on #128. Base is simonfaltum/plugin-hooks-commands; only the last commit is new here. Merge #128 first, then retarget this to main.

Why

Experimental skills are installable today only through the CLI's files overlay (databricks aitools install --experimental), so they never get the out-of-band updates the plugin channel provides, and graduating a skill to stable leaves CLI users with duplicate copies to clean up. The aitools plugin-first proposal moves experimental content to its own plugin in this repo's marketplace so the agent keeps it updated and graduation becomes a pure file move.

The plugin is deliberately skills-only, and that is a policy, not a gap: Claude Code runs hooks from every enabled plugin with no shadowing or precedence, so any hook shipped here would stack with the stable plugin's hooks (#128) on every prompt, session start, and tool call. The stable databricks plugin is the platform layer (prompt router, context primer, auth hints, slash commands); databricks-experimental is a content pack that assumes the stable plugin is installed alongside it. This also means weekly-churn experimental code never runs in the always-on hook hot path.

Changes

Before: one plugin (databricks) in the marketplace; experimental skills reachable only via the CLI files overlay. Now: a second plugin, databricks-experimental, in the same self-hosted marketplace.

• .claude-plugin/marketplace.json: second entry, source: "./experimental".
• experimental/.claude-plugin/plugin.json (new): name databricks-experimental, "skills": "./" (the skill dirs sit directly at the plugin root; there is no experimental/skills/), version 0.2.0 in lockstep with the stable manifest, and a description that states the content is unstable and may change or be removed.
• scripts/skills.py: new check_experimental_plugin in validate enforces the policy: manifest exists, parses, correct name, "skills": "./" present (without it the plugin silently ships zero skills), no hooks/commands/agents/mcpServers keys and no experimental/hooks|commands|agents dirs, version lockstep with the stable manifest, marketplace source correct. Marketplace description sync now covers both plugins.
• scripts/bump_version.py: the release bump now sets the experimental manifest too, so one tag means one version across all plugin manifests.
• Docs: README (marketplace install for both plugins, updated CLI vs marketplace table), experimental/README.md (plugin install + the skills-only rule), CONTRIBUTING ("The experimental plugin is skills-only" policy section, release step update), CLAUDE.md.

Notes:

• The plugin stays on this repo's self-hosted marketplace only. It is never submitted to Anthropic's community or official listings; the stable plugin is the candidate for those.
• Skill-name collisions between skills/ and experimental/ are already rejected at manifest generation, which keeps graduation atomic. Users who update the two plugins at different times can briefly see a graduated skill twice (namespaced databricks:<name> and databricks-experimental:<name>); updating both plugins resolves it, and aitools update on the plugin channel will do that in one go (CLI-side work, tracked in the aitools plugin-first proposal).
• An explicit version field pins marketplace updates to release bumps, matching the stable plugin's release-gated model (omitting it would make every commit to main an update, which conflicts with release gating).

Test plan

• python3 scripts/skills.py validate passes
• Negative tests: validator rejects an experimental/hooks/ dir and a version skew between the two manifests (verified manually, then reverted)
• python3 scripts/bump_version.py v0.2.0 smoke test: reports all three manifests, regenerates manifest.json as a no-op (cursor side effect reverted; the release workflow owns that bump)
• claude plugin validate passes on both the plugin manifest and the marketplace manifest
• Real install in an isolated CLAUDE_CONFIG_DIR: claude plugin marketplace add <repo> + claude plugin install databricks-experimental@databricks-agent-skills run headlessly with no prompts; claude plugin details shows all 17 experimental skills, 0 hooks, 0 commands, ~1.7k always-on tokens
• Stable plugin installed alongside in the same config: inventory unchanged (skills + 2 commands + 3 hooks, no experimental leakage), both plugins coexist
• All hook test suites still pass (router 16, context 11, auth-helper 8)

This pull request and its description were written by Isaac.

[dustinvannoy-db]

LGTM