diff --git a/.forgejo/workflows/ci.yml b/.forgejo/workflows/ci.yml
index 25b6eab..af9973a 100644
--- a/.forgejo/workflows/ci.yml
+++ b/.forgejo/workflows/ci.yml
@@ -1,13 +1,16 @@
 # CI runs first; build only proceeds if all checks pass.
 #
-# Push to dev:       typecheck + lint + test → build :dev + :<sha>
-# Tag v* (release):  typecheck + lint + test → build :latest + :<sha> + :<version>
-# Pull request:      no CI run — code is validated on dev push before any PR is opened
+# Push to any branch: typecheck + lint + test
+# Push to dev:        gates + build :dev + :<sha>
+# Tag v* (release):   gates + build :latest + :<sha> + :<version>
 #
 # To cut a release:
 #   Create a release via the Forgejo UI on main with a v* tag name.
 #   The tag push triggers this workflow; build job pushes :latest + :<version>.
 #
+# PRs aren't triggered on purpose — this is a solo dev→main flow, so
+# gating on branch push is already enough.
+#
 # Required secrets (repo → Settings → Secrets → Actions):
 #   REGISTRY_USER   — your Forgejo username
 #   REGISTRY_TOKEN  — Forgejo PAT with write:packages scope
@@ -15,7 +18,7 @@ name: CI & Build
 
 on:
   push:
-    branches: [dev]
+    branches: [dev, main]
     tags: ["v*"]
     paths:
       - "src/**"
@@ -28,8 +31,18 @@ on:
       - "assets/**"
       - "fable-mcp/**"
       - ".forgejo/workflows/ci.yml"
-  # pull_request trigger intentionally omitted — all changes go through dev
-  # first, where CI already runs on push. PR runs would be redundant duplication.
+
+# Cancel older runs on the same branch when a newer push lands. Tag runs
+# get their own group implicitly (refs/tags/v1.2.3 ≠ refs/heads/dev) and
+# are never cancelled, so a release build can't kill itself mid-flight.
+concurrency:
+  group: ci-${{ github.ref }}
+  cancel-in-progress: ${{ !startsWith(github.ref, 'refs/tags/') }}
+
+# Least-privilege default. Jobs that need more (build pushes to the
+# registry) upgrade explicitly.
+permissions:
+  contents: read
 
 env:
   REGISTRY: git.fabledsword.com
@@ -63,7 +76,9 @@ jobs:
       - uses: actions/checkout@v6
 
       - name: Install ruff
-        run: pip install --break-system-packages ruff
+        # pipx keeps ruff isolated instead of polluting the system
+        # site-packages (the old --break-system-packages was a smell).
+        run: pipx install ruff
 
       - name: Lint
         run: ruff check src/
@@ -76,7 +91,7 @@ jobs:
 
       # Python 3.12 is pre-installed in the runner-base image (py3.12-node22).
       - name: Cache pip wheels
-        uses: actions/cache@v3
+        uses: actions/cache@v4
         with:
           path: ~/.cache/pip
           key: pip-${{ hashFiles('pyproject.toml') }}
@@ -88,22 +103,28 @@ jobs:
       - name: Install package with dev deps
         run: |
           /opt/venv/bin/pip install --upgrade pip setuptools wheel
+          # http-ece's build-time dep on cryptography fails under PEP 517
+          # isolation on this runner image — install it without isolation
+          # first so the editable install below finds it already built.
           /opt/venv/bin/pip install --no-build-isolation http-ece
           /opt/venv/bin/pip install -e ".[dev]"
 
       - name: Run tests
-        run: /opt/venv/bin/python -m pytest tests/ -v
+        run: /opt/venv/bin/python -m pytest tests/ -q
 
   build:
     name: Build & push image
     needs: [typecheck, lint, test]
     # Build on dev branch pushes and version tag pushes only.
-    # main branch pushes run CI for safety but do not build —
+    # main branch pushes run the gates for safety but do not build —
     # the release tag (v*) is the sole trigger for a production image.
     if: |
       github.event_name == 'push' &&
       (github.ref == 'refs/heads/dev' || startsWith(github.ref, 'refs/tags/'))
     runs-on: py3.12-node22
+    permissions:
+      contents: read
+      packages: write
     steps:
       - uses: actions/checkout@v6
 
@@ -122,11 +143,14 @@ jobs:
           echo "build_version=$BUILD_VERSION" >> $GITHUB_OUTPUT
 
       - name: Free disk space
+        # Self-hosted runner housekeeping. Two-step cleanup:
+        # 1. Prune dangling containers/images globally (stops the runner
+        #    from accumulating cruft from past failed builds).
+        # 2. Trim the BuildKit layer cache to a 5GB ceiling so the pip
+        #    mount cache survives but old intermediate layers don't
+        #    accumulate indefinitely.
         run: |
-          # Remove all unused images (including old :SHA tags) and containers.
           docker system prune -af || true
-          # Keep the local BuildKit cache bounded so pip mount cache survives
-          # but stale intermediate layers don't accumulate indefinitely.
           docker builder prune --keep-storage 5g -f || true
 
       - name: Set up Docker Buildx
@@ -147,3 +171,10 @@ jobs:
           provenance: false
           tags: ${{ steps.tags.outputs.value }}
           build-args: BUILD_VERSION=${{ steps.tags.outputs.build_version }}
+          # Registry-backed layer cache. Pull from :cache to prime
+          # BuildKit, push updated layers back to :cache so the next
+          # build starts warm even if the runner's local cache was
+          # pruned. `mode=max` exports all intermediate layers, not
+          # just the final image, which is what gives the ~80% speedup.
+          cache-from: type=registry,ref=${{ env.IMAGE }}:cache
+          cache-to: type=registry,ref=${{ env.IMAGE }}:cache,mode=max