docs: strip YAML frontmatter, fix stale refs, update gitignore

Remove Basic Memory frontmatter from all public docs (not rendered on
GitHub). Fix "instantly" wording in README, remove Florence-2 reference
in ocr-rescan doc, tone down "novel contribution" in standard-prompt
research. Add agent config files and local assets to gitignore.

Author: swaynehales

.gitignore

@@ -21,14 +21,23 @@ venv/
 # Internal working docs (not for public repo)
 _docs/
 CLAUDE.md
+AGENTS.md
+CODEX.md
+GEMINI.md
 docs/plans/
+docs/publish/
 docs/interagent-somflow/
 docs/interagent-uitag/
 docs/current-state.md
 docs/uitag-launch-steps.md
 docs/f1.3-coreml-acceleration.md
 .sediment/
 
+# Agent / session artifacts
+.claude/
+commands-tracking-etc.md
+context-snap-output.md
+
 # Worktrees
 .worktrees/
 
@@ -38,6 +47,12 @@ tools/vision-detect
 # CoreML models (generated, large — root level only)
 /models/
 
+# YOLO weights (root-level copy; bundled model is uitag/models/yolo-ui.pt)
+/yolo11s.pt
+
+# Local-only assets
+docs/examples/hero-yolo-composite.png
+
 # Environment variables
 .env
 .env.*

CONTRIBUTING.md

@@ -1,9 +1,3 @@
----
-title: CONTRIBUTING
-type: note
-permalink: uitag/contributing
----
-
 # Contributing to uitag
 
 uitag exists to answer one question accurately: _what UI elements are on this screen?_ Every contribution should improve the answer — better detection coverage, fewer false positives, faster inference, or clearer documentation of what the pipeline does and where it falls short.

README.md

@@ -1,5 +1,3 @@
-
-
 # uitag
 
 [![Tests](https://github.com/swaylenhayes/uitag/actions/workflows/test.yml/badge.svg)](https://github.com/swaylenhayes/uitag/actions/workflows/test.yml)
@@ -12,7 +10,7 @@ _287 elements detected in ~3s — text labels (Apple Vision) + icons and buttons
 
 ## Why This Exists
 
-Screenshots should be machine-readable. Every button, label, and icon should have a bounding box, a label, and coordinates — instantly, on-device, under MIT license.
+Screenshots should be machine-readable. Every button, label, and icon should have a bounding box, a label, and coordinates — in seconds, on-device, under MIT license.
 
 Apple Vision's text recognition and rectangle detection runs natively on macOS and catches most text-based UI elements. But it misses icons, toolbar buttons, and visual controls that have no text label. On ScreenSpot-Pro (1,581 targets across 26 professional applications), Vision-only detection covers 57.3% of targets. The remaining 42.7% are invisible to Vision — predominantly icons.
 

docs/api.md

@@ -1,9 +1,3 @@
----
-title: api
-type: note
-permalink: uitag/docs/api
----
-
 # API Reference
 
 The pipeline produces two core types. A `Detection` represents a single UI element — its bounding box, label, confidence score, and source (Apple Vision text, Vision rectangles, YOLO, or Florence-2). `PipelineResult` collects all merged detections for an image along with dimensions and per-stage timing data. The entry point, `run_pipeline()`, takes a screenshot path, runs detection through all enabled stages, and returns a `PipelineResult`, an annotated PIL image, and a JSON manifest string.
@@ -506,4 +500,4 @@ Each entry in `elements`:
     "merge_ms": 3.3
   }
 }
-```
+```

docs/performance.md

@@ -1,9 +1,3 @@
----
-title: performance
-type: note
-permalink: uitag/docs/performance
----
-
 # Performance
 
 uitag runs the full detection pipeline in ~5 seconds on an M2 Max laptop with `--yolo`, producing ~300 detections with 90.8% coverage on ScreenSpot-Pro. Everything runs on-device — Apple Vision handles text, and the opt-in YOLO model handles icons and visual controls. The default Vision-only mode completes in ~1 second with ~150 detections. No API calls, no model downloads beyond the YOLO optional dependency.

docs/releases/v0.4.0.md

@@ -1,9 +1,3 @@
----
-title: v0.4.0
-type: note
-permalink: uitag/docs/releases/v0.4.0
----
-
 # v0.4.0 Release Notes
 
 ## New Features
@@ -101,4 +95,4 @@ uitag render screenshot.png --manifest manifest.json -o out/
 
 This release is backed by systematic experimentation documented in:
 - Experiment log: [`docs/research/ocr-rescan-experiments.md`](../research/ocr-rescan-experiments.md)
-- Full spec: available in project archives
+- Full spec: available in project archives

docs/releases/v0.4.1.md

@@ -1,9 +1,3 @@
----
-title: v0.4.1
-type: note
-permalink: uitag/releases/v0.4.1
----
-
 # v0.4.1 Release Notes
 
 Bugfix and UX polish release following v0.4.0.

docs/releases/v0.5.0.md

@@ -1,9 +1,3 @@
----
-title: v0.5.0
-type: note
-permalink: uitag/releases/v0.5.0
----
-
 # v0.5.0 Release Notes
 
 Fine-tuned YOLO detection model integrated into the pipeline. With `--yolo`, detection coverage on ScreenSpot-Pro jumps from 57.3% to 90.8% across 26 professional applications on macOS, Windows, and Linux.

docs/research.md

@@ -1,9 +1,3 @@
----
-title: research
-type: note
-permalink: uitag/docs/research
----
-
 # Research Notes
 
 > __Note (2026-03-29):__ This document covers the original model selection research (Florence-2 for non-text detection). Florence-2 has since been superseded by a fine-tuned YOLO model (`--yolo`) that achieves 90.8% detection coverage on ScreenSpot-Pro — see [Performance](performance.md) for current numbers. VLM classification (MAI-UI-2B-bf16-v2, 96.1% accuracy) is validated and planned for v0.6.0.
@@ -124,4 +118,4 @@ Florence-2 (159MB) co-hosts successfully alongside an 8B VLM (7.4GB) in 96GB uni
 - [Set-of-Mark (SoM) Prompting](https://arxiv.org/abs/2310.11441) — Yang et al., 2023
 - [Florence-2: Advancing a Unified Representation](https://arxiv.org/abs/2311.06242) — Xiao et al., 2023
 - [MLX: Machine Learning on Apple Silicon](https://github.com/ml-explore/mlx)
-- [mlx-vlm: Vision Language Models on MLX](https://github.com/Blaizzy/mlx-vlm)
+- [mlx-vlm: Vision Language Models on MLX](https://github.com/Blaizzy/mlx-vlm)

docs/research/ocr-rescan-experiments.md

@@ -1,9 +1,3 @@
----
-title: ocr-rescan-experiments
-type: note
-permalink: uitag/docs/research/ocr-rescan-experiments
----
-
 # OCR Rescan Experiments — Apple Vision Accuracy on Special Characters
 
 > Research supporting the multi-crop ensemble rescan feature shipped in v0.4.0.
@@ -120,7 +114,7 @@ The multi-crop ensemble rescan in `uitag/rescan.py`:
 - Crops low-confidence text at 5 padding values (5, 10, 15, 20, 25px)
 - Runs Apple Vision with `accurate` + `usesLanguageCorrection = false`
 - Selects the reading with the most special characters (raw reading heuristic)
-- Runs on ANE (zero GPU contention with Florence-2)
+- Runs on ANE (zero GPU contention with other inference workloads)
 
 Combined with the light mode advantage, this achieves perfect readings
 on all test elements in the Keyboard Maestro test image.
@@ -143,4 +137,4 @@ and raw results are documented below:
 
 _Test image: Keyboard Maestro macro configuration (regex trigger + variable actions)_
 _Hardware: Apple Silicon (M2 Max), macOS, Retina 144 DPI_
-_Date: 2026-03-06 through 2026-03-08_
+_Date: 2026-03-06 through 2026-03-08_