add the zero_shot method

[AnnaWegmann]

Adds the zero_shot method: styles are defined by rewrite instructions instead of example texts. It plugs into everything the previous PRs prepared — same generate(texts, style_dict, ...) interface and labeled output, its own style_bank (so styles and n select instruction styles like formal, simple, complex, caps, lowercase, ...), the shared causal LM base class (same model/precision options, shared loaded model with prompting), and _run_prompts for generation.

The bank restores the previously deleted zero-shot prompts from git history under clean style names (curation note from the #35 spec: the one-line "more simple" prompt is dropped in favour of the detailed simplification prompt). With this method, style_texts are instructions — exactly one per style (more errors, naming the style). An instruction may place the input text itself via [DOCUMENT SEGMENT]; otherwise the text is appended at the end with a fixed "output only the rewrite" ending.

The method itself is ~30 lines of logic; the rest of the diff is the instruction bank (restored data), three tests, and the README/methods docs sections.

Closes #35. Part of #13.

Written by Claude and Anna

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[AnnaWegmann]

fix comment

[AnnaWegmann]

Restructured the test: the placeholder/scaffold behavior is now tested directly on _fill_instruction with plain strings and no mock at all, and the generate test uses the same intuitive mock style as the prompting test — the fake model answers each prompt with a reply named after its style, so the assertion reads as [["formal reply", "pirate reply"]].

Written by Claude and Anna

[AnnaWegmann]

ok!