Ship deterministic test-level impact adapters

[aspix2k]

Purpose

Extend Affected from affected-module tests to the smallest runner-native test unit that can be proven affected without hiding regressions. Every adapter has a binary outcome: deterministic exact selection, or the existing full affected-module command.

Shipped

• Gradle/JUnit Platform exact test-class selection shipped in 1.8.0. It records a complete local test-class dependency map, skips proven-unchanged tasks, selects affected Jupiter/Vintage classes and falls back to the full Gradle Test task on any uncertainty.
• Maven 3.9.x with Surefire/Failsafe 3.x reached multi-class, multi-fork exact selection in 1.13.0, with independent unit and integration-test maps and one Maven invocation.
• 1.14.0 hardened the existing Cargo, Go, Node, .NET, Python, Composer, Bundler and CMake module-level adapters: one fail-fast IDE Run session per build root, bounded cached discovery, full fallback on uncertainty and a native eight-system conformance polygon.

Non-negotiable rules

• No confidence score, probabilistic model or heuristic dependency graph may suppress tests.
• Selection uses the runner's stable native unit: class, file, node, case or named test as proven by that adapter.
• Missing, stale, partial, corrupt or incompatible evidence runs the full original module command.
• Configuration, dependency, resource, generated-code and test-set ambiguity runs fully unless the adapter proves otherwise.
• Selected, skipped, failed and cancelled runs cannot replace a complete baseline.
• CI and release remain full.
• Source paths, test identifiers, dependency maps, hashes and metrics stay local.

2.0 rollout

1. Add exact Node test-file selection #51 — Node test files for explicitly detected Jest and Vitest runners.
2. Add exact pytest test-file selection #52 — stable pytest nodes or files.
3. Add exact .NET test-case selection #53 — .NET test cases or classes through native filters.
4. Add exact CTest test selection #58 — named CTest tests from CMake/CTest metadata.
5. Prove exact Go test selection #54 and Prove exact Cargo test selection #55 — prove or reject selection below the Go package and Cargo package.
6. Add exact PHPUnit test-class selection #56 and Add exact RSpec file selection #57 — PHPUnit test classes and RSpec files or stable groups.

Maven 4 remains tracked separately in #60; 2.0 prioritizes breadth beyond the JVM.

2.0 acceptance criteria

• Each child issue defines its supported runner/version boundary and smallest stable selection unit.
• A repository-owned public fixture catches every controlled mutation and proves every unsupported case runs fully.
• At least one supported mutation runs fewer tests faster end to end.
• README, plugin description, compatibility matrix and release notes describe only adapters that passed their proof gates.

[aspix2k]

Research checkpoint: native runner filters can execute known test IDs, but the current plugin has no sound test-impact graph and PSI alone cannot cover reflection, DI, generated code, resources, or runtime callbacks. The recommended 1.8.0 scope is therefore class-level shadow mode plus the typed ImpactProvider / TestExecutor boundary, while real execution still falls back to all tests in every affected module. Enabling opt-in narrowing should require representative shadow runs with no observed false negatives. Product decision pending: keep 1.8.0 shadow-first (recommended) or accept immediate local false-negative risk and rely on full CI.