#2561 docs: DETAILS.md's "Which model?" lists IQ3_S, setup's fourth size and its recommendation since 0.1.44

本文

**What it changes.** `docs/DETAILS.md`'s "Which model?" still said "All three", listed Q2_0, IQ2_XS and IQ3_XXS, and called IQ3_XXS the best; the first-start walkthrough said the size question offers "Q2_0, IQ2_XS or IQ3_XXS". Now:
- the table has an IQ3_S row: 84 GB download, ~50 GB of experts (`setup.py` `MODELS`: 83.6 / 50.3), slowest (DETAILS.md's own decode table), best of the four by the measurement in `docs/QUANTS.md` (88% the same top token as Q8_0, IQ3_XXS 84%); IQ3_XXS's quality cell becomes "great", as in MODELS.md;
- the 64 GB sentence gives the 128K context advice for both 3-bit sizes (as the context section above already says setup does) and that setup recommends IQ3_S there for a first install since 0.1.44 (`RECOMMENDED_SIZES`; a re-run keeps the installed size);
- the size question lists all four, and that Swift 1.5 has no Q2_0.

**What it leaves alone.** Every number already in the section, the Coder paragraph, and the rest of the page. MODELS.md's and the README's 64 GB rows are #2557; this PR does not depend on it.

**How it was checked.** Against `setup.py` (`MODELS`, `RECOMMENDED_SIZES`, `installed_size()`), `docs/QUANTS.md` and DETAILS.md's speed table. A sub-agent reviewer checked it; its corrections (a measured quality cell instead of "matches the full model", the 128K advice for IQ3_S too, "for a first install") are in.

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