Add shared BGE text map and verified graph provenance

.github/workflows/generate-pages.yaml

@@ -11,6 +11,15 @@ on:
     branches: [main]
     paths:
       - "docs/**"
+      - "kb/communities/**"
+      - "data/isolates/**"
+      - "data/text_map/**"
+      - "conf/text_map.yaml"
+      - "src/communitymech/text_map_*.py"
+      - "scripts/stage_text_map.py"
+      - "scripts/embedding_pipeline.py"
+      - "scripts/check_graph_receipts.py"
+      - "src/communitymech/graph_embedding_receipts.py"
       - ".github/workflows/generate-pages.yaml"
   workflow_dispatch:
 
@@ -33,11 +42,24 @@ jobs:
       - name: Checkout
         uses: actions/checkout@v4
 
+      - name: Set up Python for map validation
+        uses: actions/setup-python@v5
+        with:
+          python-version-file: .python-version
+
+      - name: Validate and stage the configured semantic text map
+        run: |
+          python -m pip install --quiet pyyaml
+          python scripts/stage_text_map.py
+
       - name: Configure Pages
         uses: actions/configure-pages@v5
         with:
           enablement: true
 
+      - name: Verify specialty graph receipts
+        run: python scripts/check_graph_receipts.py
+
       - name: Upload docs/ artifact
         uses: actions/upload-pages-artifact@v3
         with:

conf/embedding-runtime/README.md

@@ -0,0 +1,47 @@
+# Optional text-map runtime
+
+This governed Python 3.13 environment is for explicit local embedding and
+PaCMAP builds. Normal tests, record validation and Pages rendering do not
+install it. The lock pins numerical/model dependencies independently of a
+Mech's curation environment; the encoder and map receipts record actual
+installed library versions as well.
+
+From a Mech repository root, export current semantic input with its domain
+adapter, inspect it, and then explicitly run inference and projection:
+
+```bash
+uv run python scripts/text_map_inputs.py --output workspace/text-map-inputs.jsonl
+uv run python scripts/embedding_pipeline.py inspect --input workspace/text-map-inputs.jsonl
+uv run --locked --project conf/embedding-runtime python scripts/embedding_pipeline.py embed --input workspace/text-map-inputs.jsonl --cache workspace/text-map-vectors.sqlite --profile-output workspace/text-map-profile.json --device cpu
+uv run --locked --project conf/embedding-runtime python scripts/embedding_pipeline.py project --input workspace/text-map-inputs.jsonl --cache workspace/text-map-vectors.sqlite --profile workspace/text-map-profile.json --output data/text_map
+uv run python scripts/embedding_pipeline.py check --output data/text_map --input workspace/text-map-inputs.jsonl --cache workspace/text-map-vectors.sqlite
+```
+
+Run one small explicit `--limit` adapter canary first, with separate output and
+cache paths. Verify its vector count, finite coordinates, source receipt,
+rendered page and repeat-run cache reuse before a full build. A canary is not a
+complete-corpus artifact and cannot satisfy an enabled site's full-input check.
+
+The default display limit is 50,000 records selected deterministically. All
+input records must have a verified cache entry; the map reports omitted display
+records. No all-pairs similarity matrix is allocated.
+
+Model weights are downloaded only by the explicit `embed` command. A cached
+model can be used with `HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1`. There are no
+provider API calls. Old unpinned vector caches are not assigned this profile.
+
+The existing site renderer stages a validated bundle at `text-map/` only after
+the repository enables it in `conf/text_map.yaml`. Its normal Python environment
+runs this verification without importing Torch, NumPy or PaCMAP. The build must
+fail if an enabled map is absent, stale or invalid. Preflight checks the exact
+BGE model/revision, 1024 dimensions and the fleet's 512-token window. The staging
+call passes the approved generation name through `expected_bundle`, rejecting
+a different pointer or altered generation before changing site files.
+Site deployment remains the
+publication boundary. A machine interruption during directory staging may
+leave `.text-map-recovery-*` beside the destination; preserve it for recovery.
+
+The runtime has been exercised on macOS Apple Silicon. Platform-specific model
+wheels may constrain where heavy builds run; the standard-library verifier is
+independent of those wheels. Cross-platform reproducibility is not implied by
+the lock or a fixed random seed.

conf/embedding-runtime/pyproject.toml

@@ -0,0 +1,17 @@
+[project]
+name = "mech-embedding-runtime"
+version = "1.0.0"
+requires-python = ">=3.13,<3.14"
+dependencies = [
+  "sentence-transformers==6.0.0",
+  "transformers==5.17.0",
+  "torch==2.14.0",
+  "pacmap==0.9.1",
+  "numpy==2.3.5",
+  "numba==0.63.1",
+  "scikit-learn==1.8.0",
+  "PyYAML==6.0.3",
+]
+
+[tool.uv]
+package = false