Add UnitSentinel governance and dependency maintenance

.gitattributes

@@ -0,0 +1,10 @@
+* text=auto eol=lf
+
+*.gif binary
+*.png binary
+*.jpg binary
+*.jpeg binary
+*.webp binary
+*.onnx binary
+*.sqlite binary
+*.sqlite3 binary

.github/dependabot.yml

@@ -0,0 +1,41 @@
+version: 2
+updates:
+  - package-ecosystem: pip
+    directory: /
+    schedule:
+      interval: weekly
+      day: monday
+      time: "04:00"
+      timezone: Etc/UTC
+    open-pull-requests-limit: 5
+    groups:
+      python-runtime:
+        dependency-type: production
+      python-development:
+        dependency-type: development
+
+  - package-ecosystem: npm
+    directory: /tools/evidence
+    schedule:
+      interval: weekly
+      day: monday
+      time: "04:15"
+      timezone: Etc/UTC
+    open-pull-requests-limit: 5
+    groups:
+      evidence-renderer:
+        patterns:
+          - "*"
+
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: weekly
+      day: monday
+      time: "04:30"
+      timezone: Etc/UTC
+    open-pull-requests-limit: 5
+    groups:
+      actions:
+        patterns:
+          - "*"

CHANGELOG.md

@@ -0,0 +1,39 @@
+# Changelog
+
+All notable repository changes are recorded here. UnitSentinel has not
+published a tagged release; the entries below describe the current unreleased
+portfolio baseline.
+
+## [Unreleased]
+
+### Added
+
+- Exact unit and affine-value semantics backed by a content-addressed registry.
+- Closed, bounded computation-graph decoding and tracked Z3 verification.
+- Detached positive proof certificates with independent semantic replay.
+- Verified unit-annotation repair candidates that are never applied
+  automatically.
+- Training/serving comparison with explicit plans, lineage-drift detection, and
+  indeterminate outcomes for underconstrained contracts.
+- A closed-subset ONNX metadata adapter that validates and lowers static
+  contracts without executing models.
+- Real CLI captures, architecture diagrams, measured bounded-scaling evidence,
+  result examples, and three reproducible GIF workflows.
+- Python 3.11-3.14 tests, a 95% branch-coverage floor, dependency audits,
+  installed-wheel checks, reproducible evidence replay, and CodeQL.
+- Maintainer contribution guidance and private vulnerability reporting.
+
+### Security
+
+- Generated evidence candidates remain private (mode 0600) until reviewed and
+  deliberately published through Git.
+- Untrusted documents are subject to closed schemas, canonical encodings,
+  resource ceilings, and fail-closed result states.
+- Evidence pipelines reject secrets, personal data, external resources, and
+  undeclared files.
+
+### Release boundary
+
+No license or tagged release is published until Omar records a licensing
+decision. Package metadata version `0.1.0` does not represent a public GitHub
+release.

CONTRIBUTING.md

@@ -0,0 +1,76 @@
+# Contributing to UnitSentinel
+
+UnitSentinel is maintained as a fail-closed dimensional verification system.
+Changes should keep claims narrow, inputs bounded, and evidence reproducible.
+
+The repository is intentionally unlicensed while Omar selects and documents a
+licensing boundary. Public availability does not grant reuse rights. External
+contributions are not currently solicited; open an issue before preparing a
+substantial patch or contributing third-party code, models, or datasets.
+
+## Development setup
+
+Use a supported Python version (3.11 through 3.14) and install the locked
+evidence renderer without lifecycle scripts:
+
+```bash
+python -m venv .venv
+. .venv/bin/activate
+python -m pip install --no-input -e '.[dev]'
+npm --prefix tools/evidence ci --ignore-scripts
+```
+
+Run the core quality gates:
+
+```bash
+ruff check .
+ruff format --check .
+python -m mypy
+python -m coverage run --branch -m unittest discover -s tests -v
+python -m coverage report --show-missing --fail-under=95
+python -m build
+python -m pip_audit
+npm --prefix tools/evidence audit --audit-level=high
+```
+
+Run every committed evidence check:
+
+```bash
+PYTHONPATH=src python -B -m tools.evidence.onnx_evidence --check
+PYTHONPATH=src python -B -m tools.evidence.generate --check
+PYTHONPATH=src python -B -m tools.evidence.distribution_visuals --check
+npm --prefix tools/evidence run check
+PYTHONPATH=src python -B -m tools.evidence.repair_evidence --check
+npm --prefix tools/evidence run check:repair
+PYTHONPATH=src python -B -m tools.evidence.comparison_evidence --check
+PYTHONPATH=src python -B -m tools.evidence.comparison_visuals --check
+npm --prefix tools/evidence run check:comparison
+```
+
+## Contract and evidence changes
+
+- Preserve closed decoders, explicit resource limits, deterministic ordering,
+  exact arithmetic, and fail-closed outcomes.
+- Use only fixed synthetic graphs, models, captures, and benchmark fixtures.
+  Never publish production data, credentials, host paths, or personal details.
+- Do not hand-edit generated SVG, PNG, GIF, transcript, JSON evidence, or the
+  closed manifest.
+- Regenerate through the narrow recorder for the affected slice, render from
+  the recorded sources, inspect the actual assets at original size, and rebuild
+  the manifest.
+- Commit source changes, regenerated output, and provenance records together.
+- Label measured snapshots with their environment and avoid generalizing them
+  into cross-machine performance claims.
+- Keep `REPRODUCED`, compatibility, repair, and ONNX-import claims within the
+  boundaries documented in the README and format specifications.
+
+## Pull requests
+
+Keep history linear and give each commit one meaningful responsibility. Describe
+the contract effect, adversarial cases, commands run, evidence changed, and any
+new nonclaims. Maintainer commits must have Omar Ibrahim as both author and
+committer using the repository noreply address.
+
+Report vulnerabilities through a
+[private advisory](https://github.com/omar07ibrahim/units/security/advisories/new),
+not a public pull request.

MANIFEST.in

@@ -1,10 +1,15 @@
 global-exclude *
 
+include .gitattributes
+include .github/dependabot.yml
 include .github/workflows/ci.yml
 include .gitignore
 include .python-version
+include CHANGELOG.md
+include CONTRIBUTING.md
 include MANIFEST.in
 include README.md
+include SECURITY.md
 include pyproject.toml
 include requirements-distribution.txt
 

README.md

@@ -711,7 +711,7 @@ The timing snapshot changes only through the explicit
 
 ## Local quality gates
 
-The current suite contains 493 unit, integration,
+The current suite contains 494 unit, integration,
 adversarial, release, and evidence tests. CI publishes the exact branch-aware
 coverage table and enforces a 95% combined floor on every supported Python
 minor.

SECURITY.md

@@ -0,0 +1,43 @@
+# Security policy
+
+## Supported code
+
+UnitSentinel has no published release yet. Security fixes target the current
+`main` branch only.
+
+| Version | Supported |
+| --- | --- |
+| `main` | Yes |
+| Tagged releases | None published |
+
+## Report a vulnerability privately
+
+Use [GitHub private vulnerability reporting](https://github.com/omar07ibrahim/units/security/advisories/new).
+Do not disclose a suspected vulnerability in a public issue, pull request,
+terminal capture, generated diagram, or evidence record.
+
+Include the affected revision, the smallest synthetic graph or model that
+reproduces the problem, expected and observed behavior, and the security impact.
+Do not attach proprietary computation graphs, production ONNX models, training
+data, credentials, host paths, or personal information.
+
+## Security boundaries
+
+Graphs, certificates, comparison plans, result documents, and ONNX files are
+untrusted inputs. Decoders enforce closed schemas, byte and node limits,
+canonical forms, bounded solver work, and fail-closed outcomes. The ONNX adapter
+checks a closed static metadata subset and never executes a model; parsing an
+untrusted format is still part of the attack surface.
+
+A reproduced certificate proves current semantic reproduction under the
+recorded registry and limits. It does not authenticate an issuer or establish
+artifact provenance. Evidence and benchmark fixtures are synthetic and must
+remain free of secrets and personal data.
+
+## Disclosure process
+
+Reports are triaged through the private advisory. A coordinated fix should add
+an adversarial regression test, update affected contracts or evidence, pass the
+complete Python/distribution/evidence matrix, and be disclosed only after the
+fix is available. Licensing decisions are tracked separately and are not
+security reports.

tests/test_evidence.py

@@ -493,6 +493,16 @@ def test_evidence_reader_rejects_fifo_without_blocking(self) -> None:
             self.assertEqual(completed.stdout, b"")
             self.assertEqual(completed.stderr, b"")
 
+    def test_atomic_evidence_write_stays_private_until_review(self) -> None:
+        with tempfile.TemporaryDirectory(dir=ROOT) as directory:
+            target = Path(directory) / "candidate.txt"
+
+            evidence_generate._atomic_write(target, b"synthetic evidence\n")
+
+            self.assertEqual(target.read_bytes(), b"synthetic evidence\n")
+            self.assertEqual(stat.S_IMODE(target.stat().st_mode), 0o600)
+            self.assertEqual(list(target.parent.glob(".candidate.txt.*.tmp")), [])
+
     def test_renderer_binary_override_is_explicit_and_bounded(self) -> None:
         with tempfile.TemporaryDirectory(dir=ROOT) as directory:
             executable = Path(directory) / "node"

tools/evidence/generate.py

@@ -198,7 +198,7 @@ def _atomic_write(path: Path, payload: bytes) -> None:
             with suppress(OSError):
                 os.close(descriptor)
     try:
-        os.chmod(temporary, 0o644)
+        # Keep generated bytes private until review and Git publication.
         os.replace(temporary, path)
     except BaseException:
         with suppress(OSError):