Route bot questions to issues and add branch and media configuration

[d3cker]

Bot questions now pause work in the GitHub issue until an authorized reply arrives. This includes the initial analysis: asking the user to choose a proposed approach no longer starts an implementation session immediately. The dispatcher saves the question before publication, reassesses clarification replies, and retains the conversation across restarts and uncertain comment delivery. Human replies work with either a shared or separate GitHub posting account; marked bot comments never count as answers.

• Interpret base branch requests in natural language with the configured main model, validate the selected ref against the request and Git remote, and ask before creating a worktree when the choice is unclear or missing. Preserve the chosen base across retries and follow-up rounds.
• Declare text, vision, and audio capabilities. Analyze media in a separate read-only session without switching the main model; the setup wizard asks for a vision helper when needed.
• Always load the bundled Markdown bot prompt, with optional project instructions. Install the required runtime hooks in bot worktrees automatically.
• Close known task session tabs after a PR is closed or merged, including manual GitHub actions when auto-merge is disabled. Preserve session history, defer busy tabs until idle, recover missed closure events after TUI startup, and allow manual reopening through /bot.
• Document installation, feature-branch updates, model setup, and runtime behavior. Existing configuration remains valid; the package version is 0.5.0-beta.4.

Validation: npm run check passed type checking, all 95 tests, and the build. Coverage includes real Git worktrees and remote-branch checks, unanswered analysis questions, shared-account replies, uncertain delivery and restart recovery, follow-up rounds, media model isolation, manual PR closure, and TUI tab cleanup. Model responses and TUI interactions are mocked in the regression tests; the updated behavior still needs end-to-end validation in the external OpenCode 2 test installation.