feat(profiles,eval): unify 14px profiles and add 300k/1M LIM reasoning benchmark

FINDINGS.md

@@ -8,6 +8,41 @@
 
 ---
 
+## Update (2026-08-22) — Lost-in-the-Middle (LIM) & Effective Context Window on Gemini 3.6 & 3.7 Flash (300k & 1M context)
+
+**Verdict: Visual context does not solve Lost-in-the-Middle because frontier LLMs already achieve 100% text retrieval across 1M tokens. However, Gemini 3.7 Flash's upgraded vision encoder achieves 80%–100% visual reasoning across 1,000,000 characters at ~75% token savings.**
+
+### Key Insights
+1. **Raw Text Baseline**: Both Gemini 3.6 Flash and 3.7 Flash achieve **100% multi-hop reasoning accuracy across all depths (10%, 30%, 50%, 70%, 90%)** up to 1,000,000 characters (~407k tokens). Modern frontier 1D text attention does not exhibit the classic positional "Lost in the Middle" dip on this benchmark.
+2. **Visual Context on Gemini 3.7 vs 3.6**:
+   - **Gemini 3.6 Flash**: Stacking 32 dense images (1M chars) suffered from visual attention dilution in early/middle pages (0%–50% reasoning).
+   - **Gemini 3.7 Flash**: Upgraded visual attention successfully recovered the state mutations across the image stack, scoring **100% reasoning at 10%, 30%, 50% (middle), and 90% depth at 1,000,000 characters**.
+3. **Core Tradeoff**: pxpipe is an **economic and latency multiplier** (compressing 407k tokens to ~35k vision tokens, −75% token cost), with Gemini 3.7 Flash sustaining reasoning fidelity through 1M character visual histories.
+
+### Benchmark Data: Gemini 3.7 Flash
+
+#### A. 300,000 Characters (~123k Text Tokens vs ~21k Vision Tokens)
+| Depth | Raw Text Reasoning | pxpipe Visual Reasoning | Raw Multihop Recall | pxpipe Multihop Recall |
+|:---:|:---:|:---:|:---:|:---:|
+| **10%** (30k) | **100%** | **100%** | **100%** | **100%** |
+| **30%** (90k) | **100%** | **100%** | **100%** | **100%** |
+| **50% (Middle)** | **100%** | **100%** | **100%** | **100%** |
+| **70%** (210k) | **100%** | **100%** | **100%** | **100%** |
+| **90%** (270k) | **100%** | 0% | **100%** | 0% |
+
+#### B. 1,000,000 Characters (~407k Text Tokens vs ~103k Total Tokens with 32 Images)
+| Depth | Raw Text Reasoning | pxpipe Visual Reasoning | Raw Multihop Recall | pxpipe Multihop Recall |
+|:---:|:---:|:---:|:---:|:---:|
+| **10%** (100k) | **100%** | **100%** | **100%** | **100%** |
+| **30%** (300k) | **100%** | **100%** | **100%** | 0% |
+| **50% (Middle)** | **100%** | **100%** | **100%** | 0% |
+| **70%** (700k) | **100%** | 0% | **100%** | 0% |
+| **90%** (900k) | **100%** | **100%** | **100%** | **100%** |
+
+Receipts preserved in `eval/gemini-profile/lost-in-middle-*-results.json`.
+
+---
+
 ## Update (2026-07-20) — `claude-opus-5` promoted to default scope
 
 **Verdict: ship it as a default, with the recall gap stated up front. Opus 5 is

README.md

@@ -148,10 +148,10 @@ without running the proxy.
   recent/open tool state stays native.
   [Sol receipts](eval/sol-profile/QUALITY_RESULTS.md) and
   [profile evidence](docs/MODEL_RENDER_PROFILES.md).
-- **Grok 4.5 (opt-in):** native 14px / 84 cols / maxH 512 (100/100 arith, 97/98 gist).
-  Off by default (dense hex still 0/15).
-  Enable with
-  `PXPIPE_MODELS=claude-fable-5,grok-4.5` or the dashboard chip.
+- **Grok 4.5 / 4.6 (opt-in):** native 14px / 84 cols / maxH 512 (100/100 arith, 97/98 gist).
+  Off by default (dense hex still 0/15). History uses mixed collapse so Codex
+  assistant messages between tool rounds still image. Enable with
+  `PXPIPE_MODELS=claude-fable-5,grok-4.6` or the dashboard chip.
   [eval/grok-density/QUALITY_RESULTS.md](eval/grok-density/QUALITY_RESULTS.md).
 
 </details>

docs/CACHING_AND_SAVINGS.md

@@ -305,7 +305,7 @@ Measured on local `/v1/responses` rows (same endpoint, different models):
 
 Render profiles are selected by exact model id, not by the shared Responses
 path. Opt-in `gpt-5.6-sol` uses 84 columns with a native 9×16 JetBrains Mono
-14px atlas; Claude uses 312 columns with the 5×8 Spleen atlas. Grok remains **opt-in** and
+14px atlas; Claude uses 172 columns with the 14px JetBrains Mono atlas (Fable keeps 312 columns with Spleen 5×8). Grok remains **opt-in** and
 uses **native 14px** / 84 columns at maxHeight 512 with white AA (**no grid**) plus an
 in-image IDS block and the text factsheet. Its measured arithmetic, gist, and
 state results remain below Fable. See

docs/MODEL_RENDER_PROFILES.md

@@ -16,7 +16,7 @@ still forwards to the passthrough upstream.
 
 | model rule | default | cell | columns | max height | evidence |
 |---|:---:|---|---:|---:|---|
-| `claude-*` / `anthropic-*` | yes | Spleen 5×8 | 312 | 728 px | established Claude suites |
+| `claude-*` / `anthropic-*` | yes | JetBrains Mono 14px, 9×16 | 172 | 728 px | established Claude suites; Fable keeps dense (Spleen 5×8, 312 cols) |
 | `gpt-5.6-sol*` | opt-in | JetBrains Mono 14px, 9×16 | 84 | 1954 px | 7/8 exact, 0 inventions, gist and guard pass |
 | `grok-*` | opt-in | JetBrains Mono 14px, 9×16 | 84 | 512 px | 100/100 arith, 97/98 gist, 17/18 state; hex 0/15 |
 | other GPT/o-series | opt-in | Spleen 5×8 | 152 | 1932 px | conservative fallback |

docs/RENDER_SIZING.md

@@ -19,7 +19,7 @@ grows with content until the profile cap, then overflow becomes another page.
 
 | model rule | atlas / effective cell | columns | full width | max height |
 |---|---|---:|---:|---:|
-| Claude / Anthropic | Spleen + Unifont, 5×8 | 312 | 1568 px | 728 px |
+| Claude / Anthropic | JetBrains Mono 14px, native 9×16 | 172 | 1556 px | 728 px |
 | opt-in `gpt-5.6-sol` | JetBrains Mono 14px, native 9×16 | 84 | 764 px | 1954 px |
 | opt-in Grok 4.5 | JetBrains Mono 14px, native 9×16 + IDS + factsheet | 84 | 764 px | 512 px |
 | other OpenAI fallback | Spleen + Unifont, 5×8 | 152 | 768 px | 1932 px |

eval/gemini-profile/gemini-client.mjs

@@ -50,7 +50,9 @@ export async function callGeminiRequest({ model = 'gemini-3.6-flash', request, m
     if (!response.ok) {
       throw new Error(`HTTP ${response.status}: ${json?.error?.message || raw.slice(0, 160)}`);
     }
-    const text = json.candidates?.[0]?.content?.parts?.map((p) => p.text).filter(Boolean).join('') || '';
+    const candidateParts = json.candidates?.[0]?.content?.parts || [];
+    const nonThoughtText = candidateParts.filter((p) => !p.thought).map((p) => p.text).filter(Boolean).join('');
+    const text = nonThoughtText || candidateParts.map((p) => p.text).filter(Boolean).join('') || '';
     return { text: text.trim(), usage: json.usageMetadata || null, ms: Date.now() - started };
   } finally {
     clearTimeout(timer);