v0.8.0.2: Relational realism

v0.8.0.2: Relational realism

This is a realism and correctness release. The headline change is that generated data now
reconciles the way real data does: a parent summary column like customers.total_spent
actually equals the sum of that customer's orders, so the data holds up under a real
GROUP BY ... JOIN instead of falling apart on the first query. We also closed a
correctness gap in how relative curves behave on large, multi-batch runs. 633 tests pass,
0 failures.

Added

Cross-table aggregate roll-ups. Parent summary columns are now computed from the real
child rows, so they reconcile exactly. Declare a roll-up on the column:

"total_spent": {"type": "float", "rollup": {
    "from_table": "orders", "fk": "customer_id", "agg": "sum", "column": "amount"}}

Supported aggregations are sum, count, mean, max, and min. You can add an
optional where filter to aggregate only matching child rows, for example
"where": {"status": "completed"} (a single value or a list). Reconciliation is exact,
foreign keys stay intact, and results are deterministic under a fixed seed.

Zero-config roll-up inference. When a parent column name clearly names a child table,
such as num_orders or total_orders, the roll-up is inferred automatically with no
declaration. This is deliberately conservative. On ambiguous names like total_sales or
stock_count it declines rather than guess, so it produces no false positives on the
built-in domains.

Zero-inflated distributions. A new zero_inflate parameter adds a spike of structural
zeros on top of any base distribution, for columns like free-tier MRR or no-spend months:

"spend": {"type": "float", "distribution": "lognormal", "mu": 4, "sigma": 0.5,
          "zero_inflate": 0.3}

The zeros are applied after the min clamp, so a structural zero is not lifted to the
minimum. It is opt-in, and it is not auto-applied to domains that already model zeros
through their own logic.

Fixed

Relative-curve cross-batch convergence. Relative outcome curves now hold their shape
exactly regardless of batch_size. Before this fix, a table larger than the batch size
generated in several batches would drift. For example, a curve meant to make December
four times January would land closer to three and a half times at small batch sizes. The
correction now interpolates a factor for every month and tracks the actual accumulated row
counts across batches, so single-batch and multi-batch runs match.

YAML round-trip for generation features. The rollup, zero_inflate, depends_on,
mapping, formula, and inherits_curve_from settings now survive a
save_yaml_schema then load_yaml_schema cycle. They were previously dropped on load,
which silently disabled these features for any schema committed to a misata.yaml file.

Changed

Positioning. The library is now framed around outcome-conformant generation: you
declare the outcome you want, such as a revenue curve, a fraud rate, or a set of
multi-table aggregates, and Misata generates data that hits those targets exactly while
keeping referential integrity. This framing is now consistent across the README, the
package metadata, and the MCP server, which ships usage instructions so AI agents know
what Misata does and when to reach for it.

Notes

This release extends the exact-aggregate engine described in the arXiv preprint
(2606.08736, https://arxiv.org/abs/2606.08736v1) from temporal aggregate conformance
toward relational aggregate coherence across parent and child tables.

Full changelog: https://github.com/rasinmuhammed/misata/blob/main/CHANGELOG.md