fix: harden LLM profile configuration

[h3zh1]

Summary

• accept literal context-window token counts, warn below 8192 in the Web UI, and fail before provider calls when no output space remains
• add provider-specific default endpoints for OpenAI, Anthropic, DeepSeek, OpenRouter, Groq, Moonshot, Ollama, and Zhipu GLM while preserving explicit Base URLs
• reject empty models when saving or activating profiles, reuse the selected profile secret for probes, and gracefully fall back to manual model entry when model listing is unsupported
• close the settings dialog after a successful save and keep its geometry stable with an independently scrollable content area
• update the cyber-ui pointer so AOP error extension metadata survives timeline reduction
• make full builds work on platforms without the bundled static CRE2 archive
• synchronize English and Chinese documentation with the updated behavior

Verification

• make full
• go test ./... -skip ^TestE2ETerminal -count=1
• npm --prefix web/frontend run build
• npx --yes vitest@4.1.10 run packages/viewer/src/lib/aop-reducer.test.ts (25 passed)
• Playwright Config Panel suite (6 passed)
• desktop and 390x844 mobile settings-dialog inspection

Test note

The unfiltered Go suite still contains the pre-existing TestE2ETerminalOpenAndType hang, so the broad regression command excludes tests matching TestE2ETerminal.