docs(api): document literal __VAR_n__/__CODE_n__ token limitation

Text preservation protects code blocks and variable placeholders with
internal tokens, so source text literally containing the token alphabet
can be rewritten during restoration (cli-2a5q, wontfix: probe showed
256/256 token survival across 8 languages and both model generations,
and the sync validator already hard-errors on placeholder loss).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: sjsyrek

docs/API.md

@@ -241,6 +241,7 @@ Translate text directly, from stdin, from files, or entire directories. Supports
 - `--formality LEVEL` - Formality: `default`, `more`, `less`, `prefer_more`, `prefer_less`, `formal`, `informal`
 - `--model-type TYPE` - Model type: `quality_optimized`, `prefer_quality_optimized`, `latency_optimized` (no CLI default; the API server selects the model when omitted)
 - `--preserve-code` - Preserve code blocks (markdown, etc.)
+  - Known limitation: code blocks and variable placeholders (`{name}`, `%s`, etc.; always preserved) are protected internally with `__CODE_n__`/`__VAR_n__` tokens, so source text that literally contains such tokens may be rewritten during restoration.
 - `--preserve-formatting` - Preserve line breaks and whitespace formatting
 - `--split-sentences LEVEL` - Sentence splitting: `on` (default), `off`, `nonewlines`
 - `--tag-handling MODE` - XML tag handling: `xml`, `html`