feat(conversation): add runtime thinking mode selector

[GodfatherPacino]

Description

Adds a runtime thinking-mode selector to the main conversation composer.

• Shows a localized Enabled/Disabled selector when the active runtime advertises a compatible thinking option.
• Recognizes thought_level, reasoning_effort, and thinking option IDs so Claude Code, OpenCode, and other ACP-compatible CLIs work through their advertised capabilities.
• Supports the direct Aionrs GLM-5.2 integration through the companion AionCore change.
• Hides the control when a runtime does not advertise a supported option.
• Keeps desktop and mobile conversation composers aligned.
• Adds focused component, hook, state, and utility tests.

Related Issues

No linked issue.

Type of Change

• feat: New feature
• fix: Bug fix
• perf: Performance improvement
• refactor: Code change that neither fixes a bug nor adds a feature
• docs: Documentation only
• style: Formatting only
• test: Tests only
• chore: Maintenance
• ci: CI changes
• build: Build system changes

Atomic Change Check

• This PR contains one focused, independently reviewable change.

Local Checks

• bun run lint / repository lint gate
• Format check
• TypeScript type check
• i18n type generation and validation
• Full Vitest suite via just push

Runtime Verification

• Verified manually on macOS (Apple Silicon)
• Verified manually on Windows
• Verified manually on Linux

The packaged arm64 app and embedded backend start successfully, expose the GLM-5.2 selector, and preserve Enabled/Disabled runtime state. A real request against the earlier package revealed that Anthropic-compatible requests omitted the explicit disabled block, so GLM-5.2 continued thinking despite the UI state. That backend defect is fixed by iOfficeAI/aionrs#228 and wired through iOfficeAI/AionCore#628.

The corrected Apple Silicon package passed:

• Packaged backend resource contract: 17 checks
• Native backend smoke test: /health returned 200 with version 0.1.47
• Embedded backend architecture/hash verification
• Strict deep ad-hoc signature verification
• DMG checksum verification
• Broken-symlink scan of bundled managed resources

The exact Anthropic JSON projection is covered in the Aionrs regression tests. A final external model call is intentionally left to installation-time manual verification so this PR does not use contributor provider credentials implicitly.

Screenshots

Not included. The behavior is covered by focused component tests and packaged-app verification.

Additional Context

The direct Aionrs GLM-5.2 path depends on iOfficeAI/AionCore#628, which in turn currently depends on iOfficeAI/aionrs#228. Third-party ACP integrations are capability-driven and do not hard-code Claude Code or OpenCode behavior.

Self-review

• I reviewed the diff and limited the PR to the intended feature.
• New and changed user-facing text is localized.
• Focused tests cover the changed runtime behavior.