feat: ternary partition routing for LatticeIndex

[discreteds]

Summary

Generalises LatticeIndex partition routing from exact dict lookup to ternary + specificity semantics, so a lattice suite can declare wildcard/default partitions and route missing-field contexts correctly. Built on a new MatchStrategy.EXACT_KEY evaluated by an embedded ExpressionRulesEngine over a meta rules-table — routing now reuses the exact same ternary machinery as the constraint layer, so the two cannot drift.

Spec: docs/superpowers/specs/2026-07-19-ternary-partition-routing-design.md · Plan: docs/superpowers/plans/2026-07-19-ternary-partition-routing.md

What changed

• MatchStrategy.EXACT_KEY (core/constants.py, core/compiler.py) — rule-side-wildcard-only exact match: rule UNKNOWN sentinel → 0, rule == context → 1, context-side sentinel/UNKNOWN/null vs a specific key → −1 (the asymmetry vs stock EXACT, which scores 0). Bool: rule null is the wildcard.
• Missing/null/NaN context keys (accumulator/engine.py) — _extract_partition_key fills the typed NOT_SET sentinel (None for bool) instead of raising; _normalize_partition_key treats None and float NaN as missing (spec §1: backend null/NaN ≡ absent).
• Ternary routing (accumulator/lattice.py) — exact dict hit keeps the O(1) fast path; on a miss _route runs the meta-engine (unique top-specificity winner; 0 survivors → KeyError "No lattice…"; tie → AmbiguousPartitionError, a KeyError subclass so the service's partition-miss→422 handler is unchanged). Structural checks at construction: empty / duplicate / NOT_SET-bearing / keyless → ValueError.
• Batch routing — apply_batch normalises key columns up front (absent columns filled; null/NaN coalesced to the typed sentinel; bool keeps null) so it routes identically to single apply, with no row silently dropped or double-counted.
• Load-time ambiguity validation — index(lattices, validate=True, max_witnesses=1_000_000) runs an exhaustive witness-matrix check (equivalence classes = specific values + one OTHER per dim), chunked at 100k, ValueError on cap overflow (no silent sampling). apply_auto uses validate=False so the matrix isn't rerun per call.
• Persistence round-trip tests (str/int/bool wildcard sentinels survive Lattice.save/load) and CLAUDE.md docs.

Testing

804 passed, 36 skipped, 31 pre-registered xfailed · ruff clean · mypy unchanged (94, none new) · backend purity enforced (no new # allow: tag — count stays at three).

Review trail

Executed via subagent-driven development: per-task spec+quality review on every task, an Opus whole-branch review, and an independent Codex adversarial post-implementation review. Real defects found and fixed before this PR:

• Opus review: sorted(self._map) raised TypeError on a bool-keyed partition miss, masking the KeyError and breaking the KeyError→422 contract (→ key=repr).
• Codex review: apply_auto reran the full witness matrix every call; float NaN keys weren't normalised; apply_batch errored on an absent key column where single apply tolerates it. (A fourth finding — bool NOT_SET structural skip — was adjudicated as a non-issue: bool has no NOT_SET sentinel.)

🤖 Generated with Claude Code

[sonarqubecloud]

❌ The last analysis has failed.

See analysis details on SonarQube Cloud