fix(rtk): enforce dynamic user language detection in caveman prompts

[nitsuah]

Summary

Fixes language drift in open-sse/rtk/cavemanPrompts.js where the AI unexpectedly defaults to Vietnamese instead of matching the user's prompt language.

Problem

The previous static prompt instruction in open-sse/rtk/cavemanPrompts.js:

User wrote Vietnamese, reply Vietnamese. User wrote English, reply English.

caused the model to over-index on Vietnamese patterns in certain edge cases, leading to unexpected language switching even during English interactions.

# Example Output
 ● Creating worktree(add-c-txt)
 Switched to worktree on branch worktree-add-c-txt
 C:\Users\ajhar\code\nitsuah-io\.claude\worktrees\add-c-txt
 ● Write(c.txt)
 Wrote 1 line to c.txt
 b
 File c.txt đã được tạo với nội dung b.  # <-- Drifted into Vietnamese
 result: File c.txt created with content b.
 Thought for 12s, searched for 9 patterns, read 26 files

Solution

Updated the SHARED_PRESERVE_LANGUAGE instruction to enforce dynamic language detection while preserving intentional overrides:

"CRITICAL: Remember user prompt language. Reply in same language. Wenyan/classical-Chinese levels override this. Keep code/paths/commands verbatim."

Key Changes

• Dynamic Language Detection: Instructs the model to explicitly infer and match the user's prompt language dynamically rather than relying on hardcoded language pairs.
• Preserves Intentional Overrides: Maintains support for Wenyan/classical Chinese overrides where explicitly required.
• Code Integrity: Ensures code, file paths, and CLI commands remain verbatim regardless of output language.