kabudu · kabudu · Aug 1, 2026 · Aug 1, 2026
diff --git a/.github/release-notes/v0.4.2.md b/.github/release-notes/v0.4.2.md
@@ -0,0 +1,32 @@
+# QATQ v0.4.2: KV Geometry Relevance Gate
+
+QATQ v0.4.2 adds a separate research-only profiler that tests Capacity Oracle against real KV-cache geometry without changing Oracle theorem or certificate semantics.
+
+## What changed
+
+- `qatq-kv-geometry profile` reports bounded norm, cosine, angular, duplicate, spectral, token-position, and descriptive binary metrics from versioned KV captures.
+- Exact computation is used for compact groups; larger groups use deterministic pair sampling with explicit coverage, error metadata, block processing, and hard resource ceilings.
+- A reproducible llama.cpp corpus publishes 36 profiles across Qwen2.5 and Phi3 families, three prompt and context regimes, f16 and bf16, selected layers, all exported KV heads, keys and values, and three partition layouts.
+- Every result bundle and the complete corpus are SHA-256 bound and checked in CI.
+
+## Decision boundary
+
+Real post-RoPE key groups were highly correlated, with median maximum cosine similarity 0.954426, and the study found no defensible application capacity requirement. The evidence therefore freezes further Capacity Oracle theorem expansion and prioritizes live cold-page KV compression research. The profiler emits observations only and never emits an Oracle verdict.
+
+The pinned exporter did not expose matched pre-RoPE keys. External human coding-theory review remains outstanding and is not claimed.
+
+Read the [decision report](https://github.com/kabudu/qatq/blob/v0.4.2/docs/oracle/KV_GEOMETRY_RELEVANCE.md), [release evidence](https://github.com/kabudu/qatq/blob/v0.4.2/docs/PUBLIC_RELEASE_0_4_2_EVIDENCE.md), or [machine-readable corpus](https://github.com/kabudu/qatq/tree/v0.4.2/validation/geometry-v0.4.2).
+
+## Install
+
+Install all production CLIs with the release installer:
+
+```sh
+curl --proto '=https' --tlsv1.2 -LsSf https://github.com/kabudu/qatq/releases/download/v0.4.2/qatq-installer.sh | sh
+```
+
+Or install the crate with both research features enabled:
+
+```sh
+cargo install qatq --version 0.4.2 --features oracle,geometry --locked
+```
diff --git a/.github/workflows/publish-crate.yml b/.github/workflows/publish-crate.yml
@@ -6,7 +6,7 @@ on:
       expected_version:
         description: "Cargo package version to publish"
         required: true
-        default: "0.4.1"
+        default: "0.4.2"
 
 permissions:
   contents: read

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,6 +4,8 @@ All notable changes to QATQ are recorded here.
 
 ## Unreleased
 
+## 0.4.2 - 2026-08-01
+
 ### Added
 
 - Added the feature-gated `qatq-kv-geometry` research binary with versioned capture and result contracts, bounded exact or deterministic-sampled pair analysis, norm, cosine, angular, duplicate, spectral, and descriptive binary metrics.

diff --git a/Cargo.lock b/Cargo.lock
diff --git a/Cargo.toml b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "qatq"
-version = "0.4.1"
+version = "0.4.2"
 edition = "2024"
 rust-version = "1.96"
 default-run = "qatq"

diff --git a/README.md b/README.md
@@ -9,7 +9,7 @@
   <a href="LICENSE"><img alt="License: Apache-2.0" src="https://img.shields.io/badge/license-Apache--2.0-c78a55?style=flat-square"></a>
   <img alt="Rust edition 2024" src="https://img.shields.io/badge/rust%20edition-2024-547182?style=flat-square">
   <img alt="MSRV 1.96" src="https://img.shields.io/badge/MSRV-1.96-8a8f90?style=flat-square">
-  <img alt="Crate version 0.4.1" src="https://img.shields.io/badge/crate-0.4.1-2a4e66?style=flat-square">
+  <img alt="Crate version 0.4.2" src="https://img.shields.io/badge/crate-0.4.2-2a4e66?style=flat-square">
 </p>
 
 # QATQ
@@ -32,17 +32,17 @@ or translate observed KV distortion into mathematical separation automatically.
 
 ## Install
 
-Install all production CLIs from the v0.4.1 GitHub release:
+Install all production CLIs from the v0.4.2 GitHub release:
 
 ```sh
 curl --proto '=https' --tlsv1.2 -LsSf \
-  https://github.com/kabudu/qatq/releases/download/v0.4.1/qatq-installer.sh | sh
+  https://github.com/kabudu/qatq/releases/download/v0.4.2/qatq-installer.sh | sh
 ```
 
 On Windows:
 
 ```powershell
-powershell -ExecutionPolicy Bypass -c "irm https://github.com/kabudu/qatq/releases/download/v0.4.1/qatq-installer.ps1 | iex"
+powershell -ExecutionPolicy Bypass -c "irm https://github.com/kabudu/qatq/releases/download/v0.4.2/qatq-installer.ps1 | iex"
 ```
 
 Or build the codec from source:
@@ -158,6 +158,7 @@ integration matrix. The concise entry points are:
 - [production readiness](docs/PRODUCTION_READINESS.md)
 - [v0.4.0 release evidence](docs/PUBLIC_RELEASE_0_4_0_EVIDENCE.md)
 - [v0.4.1 independent reproduction evidence](docs/PUBLIC_RELEASE_0_4_1_EVIDENCE.md)
+- [v0.4.2 KV geometry release evidence](docs/PUBLIC_RELEASE_0_4_2_EVIDENCE.md)
 - [release checklist](docs/RELEASE_CHECKLIST.md)
 - [roadmap](docs/ROADMAP.md)
 

diff --git a/docs/PUBLIC_RELEASE_0_4_2_EVIDENCE.md b/docs/PUBLIC_RELEASE_0_4_2_EVIDENCE.md
@@ -0,0 +1,29 @@
+# QATQ 0.4.2 Release Evidence
+
+QATQ 0.4.2 packages the KV Geometry Relevance Gate as a separate, research-only production binary. It does not change QATQ or QATC wire formats, Capacity Oracle theorem families, certificate schemas, or checker semantics.
+
+## Shipped surface
+
+- `qatq-kv-geometry profile` consumes strict versioned capture metadata and emits bounded observations.
+- Exact pairwise analysis is used below configured thresholds; larger populations use deterministic sampling with coverage and error metadata.
+- Hard metadata, capture, scalar, vector, dimension, pair, and block ceilings prevent CLI flags from widening the compiled resource envelope.
+- Each new output directory contains six SHA-256-bound evidence files and can be checked with `qatq-kv-geometry verify`.
+- Sign and threshold binarizations are explicitly descriptive and do not claim semantic distinguishability.
+
+The capture and result contracts are documented in [`oracle/KV_GEOMETRY_CAPTURE.md`](oracle/KV_GEOMETRY_CAPTURE.md).
+
+## Real-capture corpus
+
+The published corpus contains 36 profiles from 12 pinned llama.cpp captures across two model families, three prompt and context regimes, f16 and bf16, early, middle, and late layers, all exported KV heads in those layers, keys and values, and three partition layouts.
+
+Every compact-corpus pair population was evaluated exactly. Spectral concentration is marked `APPROXIMATE` because its top eigenvalue uses deterministic power iteration. The corpus, per-bundle manifests, model and capture hashes, and deterministic policy are published in [`validation/geometry-v0.4.2`](../validation/geometry-v0.4.2).
+
+## Decision and claim boundary
+
+Post-RoPE key-group maximum cosine similarities ranged from 0.871583 to 0.998824, with median 0.954426. No credible application-level source for required state count or separation was established. The required decision report therefore freezes further theorem expansion and redirects product work toward live cold-page KV compression.
+
+The pinned exporter did not expose matched pre-RoPE keys, so no pre-RoPE comparison is claimed. External human coding-theory review remains outstanding and is not claimed. See [`oracle/KV_GEOMETRY_RELEVANCE.md`](oracle/KV_GEOMETRY_RELEVANCE.md) for the complete decision record.
+
+## Verification
+
+The release commit passes formatting, all-target all-feature compilation, the full locked test suite, two geometry CLI integration tests, the repository typography gate, independent SageMath reproduction of all 27 Capacity Oracle rows, and `scripts/verify_kv_geometry_corpus.py` over all 36 published profiles.