Expand CLI automation and maintenance workflows

[JaviChulvi]

Summary

• replace ad-hoc slash-command handling with a registry, command completion, color controls, and compact/verbose/quiet display modes
• add direct session resume/continue flows and defer session/trace persistence until a turn actually starts
• add non-interactive chulk exec with structured JSON and stable exit codes
• add offline chulk init and chulk doctor workflows
• add trace inspection and HTML export commands
• update the README and roadmap for the expanded CLI surface

Reliability and safety

The review fixes folded into their related feature commits now cover:

• ordinary prompts beginning with bare help, exit, or quit aliases still reach the model
• all configuration-backed CLI modes resolve an unset project root from the current working directory
• failed plain-text exec turns write diagnostics only to stderr
• failed or blocked exec turns return a runtime-error exit code
• approval-required status tracks actual approval callbacks, including hosted MCP requests without tool-call records
• doctor reports configuration read failures instead of raising tracebacks
• doctor rejects ignored runtime databases or traces that are already tracked by Git
• doctor validates credentials and model capabilities for the primary provider and every fallback
• doctor validates the runtime directory, SQLite store destination, and trace destination independently
• init validates all scaffold targets before writing and rejects symlinked or out-of-root paths
• trace inspect/export normalize non-UTF-8 input into structured trace errors
• trace export refuses to overwrite its JSONL source, including with --force
• CI installs the provider and build dependencies exercised by the full test suite

These issues came from deriving automation readiness from incomplete proxies and from allowing command, config, path, decoding, or workflow dependency failures to escape their documented contracts.

Impact

Scripts and CI can reliably distinguish success, runtime failure, configuration failure, and approval requirements. Interactive users get quieter default progress, explicit display controls, startup session resume, and no empty session artifacts. Project initialization and diagnostics remain offline and project-scoped.

Validation

• python -m pytest — 353 passed in a fresh Python 3.13 environment
• clean no-isolation wheel build
• python -m compileall chulk
• ruff check .
• git diff --check origin/main...HEAD