feat: add dataset grain and aggregate reconciliation rules

Goal and scope

Declare composite row grain and aggregate conservation through Phlo's existing neutral quality API. Second PR in the stack; its base contains explicit event-time freshness.

Change map and risk

• Add public grain() and reconcile_aggregate() factories, translated by Pandera into bounded SQL checks.
• Support independent equality filters, existing partition scopes, runtime-selected logical relations, and decimal-exact scalar comparison.
• Preserve Dataset severity, evidence, and publication policy. Reject pinned catalog references and conflicting evidence names.
• Combine strict freshness and dataset-constraint handling for empty data while retaining the legacy skip for other checks.
• Add a disposable real-lakehouse demonstration and author documentation.

Original verification

• Original head: dataset constraints, public exports, and neutral-provider suites: 59 passed, including mixed explicit-freshness/constraint execution.
• Identical before/after procedure on Trino 477, Nessie 0.108.3, Iceberg/MinIO and PostgreSQL: a grouped join inflates settled revenue 41 → 52 while grain/range checks pass. Baseline publishes; reconciliation reports delta 11 and blocker policy retains durable draft state. Warning policy publishes without changing the failed quality verdict.
• Repaired candidate source/target totals 54, while main remains 41, proving both relations follow the candidate ref. This does not prove physical Nessie promotion or cross-table snapshot locking.
• Reproduce: uv run --locked python examples/lakehouses/wap-failure-lab/scripts/constraints_demo.py --output /tmp/constraints.json. Docker and make setup are required; the script provisions disposable services. See the WAP Failure Lab README for declarations.
• Full integrated baseline: 7,417 passed; core regression: 447; real storage regression: 101; installed-wheel verification: 40 packages. These broader checks cover the final stack.
• Links and generated references pass. Final combined production documentation rendering exceeded the development orb's memory limit.
• New feature: regression coverage includes composite/null keys, scoped aggregates, empty data, tolerance boundaries, evidence-name collisions, and interaction with explicit freshness.

Review corrections

• Legacy checks remain skipped on empty data even alongside grain or reconciliation checks. Explicit freshness and dataset constraints still execute.
• Accepted-values checks participate in collision validation using their emitted evidence name, not their display name.
• Schema and SQL checks select the same runtime Trino ref through the shared resolver. Opposite main/candidate contents reproduce the mismatch before the fix and agree after it. The regression uses real DuckDB SQL behind the Trino driver interface, not a live Trino server.
• All 269 selected Pandera tests pass. The original head also passed GitHub CI, including documentation rendering.

Post-correction integrated stack: make check passes with 7,419 tests, four skipped and 194 deselected; make test-core-regression passes all 447 tests. These are combined-stack results, not a separate full baseline at this isolated head.

Recovery and maintainer decisions

Opt-in; no migration or deployment. Remove constraint declarations before reverting their implementation. A source revert does not undo publication. No source dependency inference or private-provider compatibility claim. Maintainer review is pending; no independent reviewer or release qualification is claimed.