feat: evidence policy, strict comparator, Snowflake recorder and append-only guard (#3)

* feat: evidence policy file and strict comparator

* feat: cases, pending files, DuckDB replay, verify and audit

* refactor: evidence and recorder layout

* refactor: evidence crates follow the Rust structure rules

* feat: Snowflake recorder, append-only guard and CI wiring

* docs: evidence workflow, recorder settings and integrity

* test: neutral stand-in column name

* fix: rerun a batch case by case when Snowflake wraps a statement error

* fix: install the connector's pandas extra for Arrow fetches and pin pyarrow 23

* fix: store only a hash of the query ID and no warehouse error messages

* evidence: first Snowflake recording of the seed cases

Author: kvlonge

.github/workflows/ci.yml

@@ -21,8 +21,17 @@ jobs:
     env:
       CARGO_TERM_COLOR: always
       CARGO_INCREMENTAL: "0"
+      # The Snowflake recorder is never built or run in CI.
+      RECORDER: "0"
+      # `uv run` would otherwise sync the default groups, recorder included.
+      UV_NO_SYNC: "1"
+      # The append-only guard compares recorded results with the pull request base or the
+      # previous main commit.
+      EVIDENCE_BASE: ${{ github.event.pull_request.base.sha || github.event.before || 'origin/main' }}
     steps:
       - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
       - uses: astral-sh/setup-uv@v5
         with:
           version: "0.12.3"
@@ -36,8 +45,8 @@ jobs:
           shared-key: verify
       - name: Install Python
         run: uv python install 3.12
-      - name: Install development tools
-        run: uv sync --locked
+      - name: Install development tools without the recorder
+        run: uv sync --locked --no-group recorder
       - name: Run static checks
         run: make check-ci
       - name: Run tests

.gitignore

@@ -29,3 +29,7 @@ __pycache__/
 /.agents/skills/
 /.claude/skills/
 /.opencode/skills/
+
+# Derived evidence work files
+/evidence/pending/
+/evidence/audit/

AGENTS.md

@@ -25,6 +25,9 @@
 - Everything that could be weakened to make a failing case pass (comparator tolerances, required
   edge-case inputs, known divergences) lives in one reviewed policy file. There are no per-case or
   per-rule overrides elsewhere.
+- Tests that need a result use DuckDB stand-ins written to temporary directories and marked as
+  test fixtures. Never commit a stand-in under `evidence/results`; the evidence guard rejects any
+  result the recorder did not write.
 - When a case fails, either fix the rule or make it refuse. Never widen a rule to raise the
   conversion rate.