feat: accumulator engine — Build/Apply for maximal consistent rule combinations

[discreteds]

Summary

• New engine: AccumulatorEngine with explicit build() → Lattice and apply(lattice, context) → AccumulatorResult two-phase API. Builds the lattice of maximal consistent rule combinations with accumulated numerics (the "accumulator builds; the filter ranks" pattern from the architecture analysis).
• Shared metadata extension: DimensionRole enum (CONSTRAINT / CONTEXT_KEY) on Dimension — backward-compatible, filter engine ignores it.
• AccumulatorCompiler: Compiles coalesce, compatible, and NA flag expressions for EXACT, RANGE, GREATER_THAN, LESS_THAN strategies. Sentinel-safe conditional coalesce prevents corruption from greatest/least on sentinel values.
• Prime-product combination DNA: Each rule assigned a unique prime per partition. Combination identity is the product of member primes. Frontier filter uses prime divisibility for Pareto-dominant subset detection. Checked multiplication raises OverflowError at int64 boundary.
• 128 new tests: unit tests for all components, cross-backend parametrized apply tests (7 backends), edge cases (empty rules, multiple aggregates, frontier filter isolation, range/GT/LT integration), and performance benchmarks (build scaling by rules/dims/density, apply scaling).
• Backlog captured: Multi-lattice composition discussion doc for cross-authority lattice joins.

Architecture

The accumulator engine answers "what is the maximal consistent combination of rules, and what numeric does it accumulate?" — complementing the existing filter engine which answers "which rules apply to this context?"

Build phase: Python-controlled breadth-first iteration where each step's work (cross-join, compatibility filter, dimension coalesce, numeric accumulation) runs as a relation() pipeline via mountainash.expressions. Frontier filter removes dominated combinations via prime-product divisibility anti-join.

Apply phase: reuses ExpressionRulesEngine by remapping co_ prefixed lattice columns to filter-engine-compatible DimensionsMetadata.

New Files

File	Purpose
accumulator_engine.py	AccumulatorEngine — build, build_all, apply, apply_auto
accumulator_compiler.py	Coalesce/compatible/NA flag expression compilation
accumulator_result.py	AccumulatorResult(RuleResult) with accumulator accessors
aggregate.py	Aggregate pydantic model
lattice.py	Lattice data class
primes.py	Prime sieve, allocation, checked multiply

Test plan

• 618 tests passing (128 new), 0 failures
• Cross-backend apply phase verified across all 7 backends (3 xfailed on ibis-polars #78)
• Worked example from architecture analysis reproduced end-to-end (3 rules → 6 outermost rulesets)
• Edge cases: empty rules, multiple aggregates, unsupported operations, frontier filter isolation, sentinel coalesce
• Performance benchmarks: build scales with unknown density (30ms at 10%, 215ms at 70% for 10 rules/4 dims); apply is ~4ms constant
• Full existing test suite green — no regressions

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (develop@db960b8). Learn more about missing BASE report.
⚠️ Report is 50 commits behind head on develop.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop      #42   +/-   ##
==========================================
  Coverage           ?   93.90%           
==========================================
  Files              ?       13           
  Lines              ?      623           
  Branches           ?       89           
==========================================
  Hits               ?      585           
  Misses             ?       28           
  Partials           ?       10           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[sonarqubecloud]

❌ The last analysis has failed.

See analysis details on SonarQube Cloud