fix(prompt): stop base system prompt forcing language guesses on all providers

[quiet-node]

Description

Fixes #349. Since v0.16.0 the base system prompt carried an always-on instruction to reply in the language of the user's latest message, with Vietnamese and Japanese named as examples. It rode in on the search language-parity work (#326) but reaches every provider on every chat turn, and on small local models the named-language tokens can steer replies into Vietnamese even for English questions. This PR removes that instruction from the base prompt.

Key changes

• Remove the two-line language-matching instruction from src-tauri/prompts/system_prompt.txt, the exact inverse of the hunk feat: search language parity across retrieval and answers #326 added there.
• Built-in web search keeps its language parity: the scoped rule in the search writer prompt, anchored to the per-message resolved language, is untouched.

Testing

• bun run test:all:coverage: frontend and backend green, 100% coverage gate holds.
• bun run validate-build: clean.

Note: existing installs pick up the corrected default automatically; the config loader replaces a non-customized prompt.system with the current built-in prompt on load.