v0.0.26: provider profiles and capability-aware reasoning

[ackness]

Summary

• introduce independent provider connection profiles, key namespaces, and automatic migration of legacy presets
• expose reasoning controls from actual model capabilities and translate them safely for OpenAI, Anthropic, DeepSeek, and Qwen-compatible providers
• reorganize LLM settings around connections, models, slots, and advanced options while clearing stale overrides
• classify long-lived user/developer documentation under docs/ and move implementation-only material out of the published docs tree
• bump all workspace package versions and public release metadata to v0.0.26

Compatibility fixes

• omit tool_choice for DeepSeek V4 thinking requests and keep tool-call assistant content non-null
• avoid sending Anthropic effort fields to Claude 4.0 models
• restrict gpt-5-pro reasoning effort to high
• prevent disabling thinking on Qwen thinking-only models
• preserve distinct credentials and model references when multiple connections share one provider family

Validation

• pnpm release:preflight
• pnpm peers check
• pnpm lint
• pnpm deps:check
• pnpm check:i18n
• pnpm test (46 tasks; web 533, store 719, runtime 860, server 716)
• pnpm --filter @covel/desktop build (web/main/preload build, server staging, native addon rebuild, and staged server smoke tests)
• pnpm exec prettier --check ...
• git diff --check

Release plan

After this PR and the main CI run pass, merge it and tag the merge commit as v0.0.26. The tag triggers the Build Desktop workflow and publishes the GitHub release artifacts.

The current release workflow intentionally produces unsigned macOS and Windows artifacts because signing and notarization secrets are not configured.