Add practical context windows section — usability not ceilings

VRAM must hold weights AND KV cache. What's left after weights is
your real context window. For a 10K-line project (~100-150K tokens):
- 1 card: 32-100K usable, file-by-file workflow
- 2 cards: 80-180K usable, whole-project-in-one-shot workflow
- This Opus session: 1M tokens (neither setup comes close)

Comparison table vs Claude Opus 1M session for perspective.
Keep Pro for the hard stuff, use local for the daily 80%.

https://claude.ai/code/session_01PtYTPherSJaxDEVPgF6Nxu

docs/gpu-setup-research.md

@@ -252,6 +252,50 @@ Think of it this way:
 - **1 card**: B+ coding (MoE or small dense models, squeezed context)
 - **2 cards**: **A- coding** (full dense 27B/32B models, comfortable context, higher quant options)
 
+### Practical Context Windows (Usability, Not Ceilings)
+
+Context window "support" is a ceiling, not what you actually get. VRAM must hold both the
+model weights AND the KV cache. What's left after weights determines your real context.
+Quality also degrades toward the edges of a model's context window.
+
+**Reference: A 10,000-line codebase ≈ 100-150K tokens** (varies by language/comments).
+This Claude Opus session uses a **1 million token** context window for comparison.
+
+#### 1 Card (16GB) — Practical
+
+| Model | Weights | Free for KV | **Usable context** | 10K-line project? |
+|-------|---------|-------------|-------------------|-------------------|
+| Qwen3.5-35B-A3B (MoE) | ~12GB | ~3GB | **32-50K tokens** | **No — ~1/3 of it** |
+| Qwen3.5-9B (dense) | ~6GB | ~9GB | **80-100K tokens** | **Mostly — but B+ quality** |
+| Qwen2.5-Coder-14B | ~10GB | ~5GB | **16-24K tokens** | **No — a few files at a time** |
+
+**Workflow on 1 card:** You're feeding files in chunks. Good for "fix this function" or
+"explain this file." Not for "read my whole project and refactor the auth system."
+
+#### 2 Cards (32GB via NVLink) — Practical
+
+| Model | Weights | Free for KV | **Usable context** | 10K-line project? |
+|-------|---------|-------------|-------------------|-------------------|
+| **Qwen3.5-27B (dense)** | ~17GB | ~14GB | **80-128K tokens** | **Yes — most/all of it at A-** |
+| Qwen3.5-35B-A3B (MoE) | ~12GB | ~19GB | **128-180K tokens** | **Yes with room to spare (B+)** |
+| Qwen2.5-Coder-32B | ~20GB | ~11GB | **32-48K tokens** | **Partial — but strong A- on what it sees** |
+
+**Workflow on 2 cards:** You can dump most/all of a 10K-line project in one shot with the
+27B dense model. That's the real workflow change — "here's my whole project, find the bug"
+becomes possible locally.
+
+#### vs This Claude Session
+
+| Setup | Usable context | vs Opus 1M | Whole-project workflow? |
+|-------|---------------|------------|----------------------|
+| 1x RTX 5000 (best) | ~50-100K | 5-10% | No — file by file |
+| **2x RTX 5000 (best)** | **~128-180K** | **13-18%** | **Yes — for 10K-line projects** |
+| Claude Opus (this session) | 1,000K | 100% | Yes — for anything |
+
+**Neither setup replaces this session** for complex multi-file work across a 50K+ line
+codebase. That's why you keep Pro. But 2 cards handles the daily "read my project and
+help me code" workflow locally with no rate limits — and that's 80% of the work.
+
 ### Estimated Inference Speed
 
 | Model | 1x RTX 5000 | 2x RTX 5000 (NVLink) |