SwiftInferProperties

Type-directed property inference for Swift. Surfaces idempotence, round-trip pairs, and algebraic-structure (semigroup / monoid / group / semilattice / ring) candidates from function signatures, cross-function pairs, and existing unit tests — for human review, never silent execution.

Status: v1.41.0 (thirty-eighth calibration cycle). Closes the v1.35 cycle-32 finding: RefactorClusterAnalyzer.classify now uses a two-layer dominant-pattern rule. OrderedSet's 29-suggestion cluster reclassifies from the misleading algebraicStructure (only 14% algebraic — fired under the pre-v1.41 "any 2 distinct templates wins" rule) to dual-style-consistency cluster (dual-style 12 entries is the dominant single template). The curated suggestion text now correctly points at SetAlgebra conformance. Layer 1: algebraic-collective dominance (2+ distinct algebraic templates AND their sum ≥50% of total → algebraicStructure). Layer 2: single-template most-numerous wins (with the pre-v1.41 priority order retained as tie-breaker). Layer 3: ≥4 total → generalCluster catch-all. End-to-end verified: only OrderedSet changes classification on OrderedCollections; ComplexModule stays algebraicStructure (12/20 = 60%); the other 6 OC clusters had genuine algebraic dominance (57–67%) and are unchanged. Constraint Engine refactor (v1.36–v1.40) untouched — v1.41 modifies only the cluster-classification layer. Test count 2097 → 2103 (+6). v1.42+ priority: user's call from higher-order property composition (PRD §20.2 lookahead), backlog items (cross-type abstraction discovery, incremental indexing, NL query DSL, SQLite backend), or the still-deferred test-execution evidence architectural shift. The full design lives in docs/SwiftInferProperties PRD v1.0.md and the v1.41 plan at docs/v1.41 Calibration Plan.md. Current performance baseline: docs/perf-baseline-v1.41.md. [previous: v1.40.0] (twenty-third calibration cycle — fifth empirical-only release after cycles 6 = 26.7%, 14 = 34.8%, 17 = 52.3%, 20 = 48.8%; binary-equivalent to v1.25.0). Headline: 25/37 = 67.6% Possible-tier acceptance rate — outcome A (Aggregate ≥ 60%); +18.8pp from cycle-20's 48.8% (the largest single-cycle aggregate jump in the loop's history). §19 ≥70% target now within +2.4pp — sample-noise band on n=40. Five-point trajectory: 26.7% → 34.8% → 52.3% → 48.8% → 67.6%. The cycle-20 non-monotonic step (-3.5pp) is validated as calibration-trade-off + sample-shift — the surviving v1.25 surface composition has materially higher per-template accept rates (round-trip 85.7%, dual-style-consistency 100%, idempotence-lifted 66.7%). Drivers: V1.21.C + V1.22.B/D + V1.24.A closed cross-product round-trip noise; V1.24.B + V1.24.C + V1.25.A closed direction-op + non-deterministic lifted-idempotence rejects; V1.24.D + V1.25.A reduced idempotence non-lifted from 23 picks (5-cycle 0%) to 3 picks (all unknown) — 0% drag eliminated. V1.18.C dual-style 100% rate-stability across 3 consecutive measurement points (largest mechanism-class precision contribution in loop history). Cycle-24 priority list rotated post-v1.26 (top: FP approximate-equality template arm 10-cycle carry-forward; 2 NEW cycle-23 findings — Algo Iterator-like survivors + OC bucket/word inverse-pair). The full design lives in docs/SwiftInferProperties PRD v1.0.md and the v1.26 plan at docs/v1.26 Calibration Plan.md. Current performance baseline: docs/perf-baseline-v1.26.md (v1.25 carry-forward; v1.25 + v1.24 + earlier baselines retained for forensic comparison).

What it does

swift-infer discover --target Foo walks Sources/Foo/, runs the M1–M8 template engine + the M1.4 cross-function pairing pass + the TestLifter M1–M16 test-side cross-validation (round-trip / idempotence / commutativity / monotonicity / count-invariance / reduce-equivalence detectors, plus Int/String/Bool/Double inferred preconditions, inferred domains for both round-trip pairs and general consumer-producer chains, and predicate equivalence-class advisories with multi-marker partitions + N-class same-target enum exhaustiveness), and emits ranked suggestions per PRD §4. Each suggestion ships an explainability block with both "why suggested" and "why this might be wrong" (PRD §4.5). Strong-tier suggestions can be triaged with --interactive; accepted suggestions emit property-test stubs to Tests/Generated/SwiftInfer/<TypeName>/ and conformance stubs to Tests/Generated/SwiftInferRefactors/<TypeName>/ (PRD §16 #1 — never edits existing source).

swift-infer drift --target Foo diffs current discovery output against a .swiftinfer/baseline.json snapshot and emits non-fatal CI-friendly warnings for new Strong-tier suggestions that lack a recorded decision (PRD §9). Drift never fails the build.

swift-infer metrics aggregates .swiftinfer/decisions.json files into per-template acceptance / rejection / suppression rates (PRD §17.2). Useful for the empirical-tuning loop SwiftInfer's signal weights are calibrated against — see docs/calibration-cycle-1-findings.md, docs/calibration-cycle-2-findings.md, docs/calibration-cycle-3-findings.md, docs/calibration-cycle-4-findings.md, docs/calibration-cycle-5-findings.md, docs/calibration-cycle-6-findings.md, docs/calibration-cycle-7-findings.md, docs/calibration-cycle-8-findings.md, docs/calibration-cycle-9-findings.md, docs/calibration-cycle-11-findings.md, docs/calibration-cycle-12-findings.md, docs/calibration-cycle-13-findings.md, docs/calibration-cycle-14-findings.md, docs/calibration-cycle-15-findings.md, docs/calibration-cycle-16-findings.md, docs/calibration-cycle-17-findings.md, docs/calibration-cycle-18-findings.md, docs/calibration-cycle-19-findings.md, docs/calibration-cycle-20-findings.md, docs/calibration-cycle-21-findings.md, docs/calibration-cycle-22-findings.md, docs/calibration-cycle-23-findings.md, docs/calibration-cycle-24-findings.md, docs/calibration-cycle-25-findings.md, docs/calibration-cycle-26-findings.md, docs/calibration-cycle-27-findings.md, docs/calibration-cycle-28-findings.md, docs/calibration-cycle-29-findings.md, docs/calibration-cycle-30-findings.md, docs/calibration-cycle-31-findings.md, docs/calibration-cycle-32-findings.md, docs/calibration-cycle-33-findings.md, docs/calibration-cycle-34-findings.md, docs/calibration-cycle-35-findings.md, docs/calibration-cycle-36-findings.md, docs/calibration-cycle-37-findings.md, and docs/calibration-cycle-38-findings.md for cycle-1 through cycle-38 results (cycle 10 was the v1.13 zero-behavior-change hoist refactor; no findings doc).

Relationship to SwiftPropertyLaws

SwiftInfer is a one-way downstream of SwiftPropertyLaws:

SwiftInferProperties → SwiftPropertyLaws (PropertyBackend, DerivationStrategist) → swift-property-based

Where SwiftPropertyLaws verifies the laws of declared protocol conformances, SwiftInfer surfaces implicit properties — and, when enough algebraic evidence accumulates on a type, suggests the standard-library or kit-defined protocol the type could conform to so SwiftPropertyLaws keeps verifying the laws on every CI run thereafter (RefactorBridge). v1.1 emits conformance proposals against Equatable, Comparable, Numeric, SetAlgebra, and (kit-defined) Semigroup, Monoid, CommutativeMonoid, Group, Semilattice.

Add to your project

.package(url: "https://github.com/Joseph-Cursio/SwiftInferProperties", from: "1.41.0")

Or run as a one-off against an existing target:

swift run swift-infer discover --target Foo
swift run swift-infer drift --target Foo

Build & test

swift package clean && swift test

The test suite is 1200+ tests across 180+ suites covering every shipped template, every TestLifter detector, the §13 performance budgets, and the §16 hard guarantees. SwiftLint config lives at .swiftlint.yml; swiftlint lint --quiet should be silent.

License

MIT — see LICENSE.