feat: add tool availability and standalone task review

[MelodyVAR]

Headless callers can advertise tools that their policy cannot use, while task self-review currently requires a Git-commit workflow. This adds extension-owned unavailable-tool declarations and lets --completion-review run once in the same invocation without requiring Git.

• Keep discovery, active schemas and generated prompts consistent, preserving explicit tool selections across availability changes and reloads.
• Preserve caller-provided input and terminal failure handling; defer automatic feedback after an unapproved external plan handoff.
• Add plan snapshot hashes, decision ownership and exit-state metadata while retaining interactive approval behavior.
• Add an opt-in runtime-info example extension with bounded local diagnostics, separate model/request token limits, and no payload or credential dumps.
• Avoid persisting an already-enabled retry preference during startup; the existing CLI configuration-preservation regression also failed on the clean base and now passes unchanged.

Validation: npm run check, pnpm run build:offline, and ./test.sh passed with Node 22.19.0 / pnpm 9.15.9. The full run includes 4,553 coding-agent tests and 672 CLI tests passing. A configured CLI inspection smoke also confirmed zero HTTP requests and unchanged configuration. Tests use isolated resources without model credentials.

Review and diagnostics remain opt-in. External approval policy and execution-time permission checks remain with their existing owners; self-review is not a correctness score.