fix(cmind): separate AI execution authority from repository config and align workflows

[Qingtao Li (QingtaoLi1)]

Summary

Separate repository-provided AI recommendations from execution authority, add shell-free Windows installation adapters, and align onboarding, migration guidance, and command templates with the new behavior.

Implementation

• Treat recommended_provider and validated legacy ai_provider / ai_cli_cmd values as repository hints only.
• Require an explicit caller choice, trusted process/CI environment, or user-local provider selection bound to the canonical workspace path.
• Save local selection only after successful setup. Updates without --ai preserve existing selection and never infer execution consent.
• Construct process arguments from a closed provider table; reject arbitrary commands, paths, additional arguments, and malformed configuration.
• Support native Windows executables and recognized Claude/Copilot npm entry points without executing shell wrappers.
• Make Git hooks opt-in and synchronization-only; remove background AI updates and reconcile recognized legacy hook blocks.
• Remove default provider permission-bypass flags.
• Preserve the newer main branch's Windows UTF-8, process-tree cleanup, and diagnostic agent-detection behavior without allowing detection to authorize execution.
• Add isolated regressions and release gates for source, installed-wheel, and Release ZIP checks.

Documentation and command templates

• Update the root README and all five CoderMind language guides.
• Document explicit initialization, local selection, migration, Windows compatibility, and opt-in hooks.
• Correct storage-layout, CLI help, and permission messaging.
• Update all 15 command templates to surface configuration/authentication/access/approval blockers and pause rather than automatically changing settings or retrying.
• Clarify that independent Release ZIP templates can still contain broader Claude permissions; removing CLI bypass flags does not revoke existing grants.

Compatibility

• New clones, moved workspaces, and new users require explicit local selection or trusted process/CI configuration.
• Valid repository hints may remain tracked but no longer authorize execution.
• Arbitrary custom commands and script wrappers are unsupported.
• AI-driven graph updates require explicit invocation; normal provider approvals may interrupt unattended workflows.

Validation

• Focused isolated security and template regressions: 812 passed, 41 skipped, 5 deselected.
• Markdown structure, template frontmatter, new relative links/anchors, and whitespace checks passed.
• AST comparison confirmed the documentation follow-up changed CLI text/comments only.
• No real AI CLI was installed or executed; developer-machine settings and actual Git hooks were unchanged.

An additional planner-test run exposed an existing assertion mismatch: test_warning_is_treated_as_incomplete expects type=warning, while the unchanged implementation returns a more descriptive warning reason. This unrelated issue was not modified.

Remaining validation

• Run remote Linux/Windows CI and installed-artifact checks.
• Verify real supported CLI installations in an isolated environment.
• Complete the original-report retest before considering the incident resolved.

No enough permission