Command inspector implementation using shared completion knowledge #242

Implement the safe v1 command inspector designed in #84 on the existing v0.8 completion stack. Research remains in #84; this child tracks implementation.

Acceptance:

• Reuse CompletionCandidate metadata and the existing lexer; share authored local facts with completion.
• Determine command/subcommand/flag/argument at the grapheme cursor without executing input.
• Display title/type, concise description and optional known usage; show truthful unavailable context.
• Keyboard palette toggle; hide during execution, panel takeover and paste atoms.
• ScreenPlan places above Bottom, below Top and alongside the scrolling Flow composer, independently of Normal/Chat.
• Bounded plain presentation at narrow widths, Safe glyphs, NO_COLOR and deterministic mode.
• No network, arbitrary help, per-keystroke programs, PTY or stored output mutation.
• Automated context/provenance/execution-boundary/layout tests; later physical host QA.

Implementation: #244. Milestone #8. Keep open pending integration and physical QA. Man/whatis/help extraction and documentation browsing are deferred.