APG v0.4 needs Go testing capability. Three distinct concerns appear in real Go repositories and they do not have the same owner.
Native Go testing is universal. Every Go repository with tests uses the standard library’s test type, its execution lifecycle, its subtests, its cleanup and isolation helpers, its parallelism rules, and its selection and caching behavior. None of that depends on a third-party dependency.
matryer/is and google/go-cmp are optional and independently selected. A
repository may use neither, either, or both. They solve different problems.
matryer/is is a deliberately minimal assertion surface: calibration release
v1.4.1 exposes four assertion families and no option surface. google/go-cmp is
a configurable semantic comparison engine: calibration release v0.7.0 exposes
an extensive option, comparer, transformer, and filter surface whose composition
determines what a test can still detect.
go-language-profile already owns general Go semantics and does not address
test-specific judgment. APG’s process skills already own implementation,
planning, debugging, and review. Neither addresses which testing owner holds
which part of a test’s contract.
The design question is how many owners APG should create, and whether a composition owner is justified. A single monolithic Go testing skill would bind universal native semantics to two optional dependencies. Component profiles alone would leave the recurring cross-component question — which owner holds which part of the contract — unowned. This ADR proposes the split and the thin composition owner, and records the alternatives considered.
APG35 authored the candidate leaves but did not integrate or test them. APG36 then performed fresh independent source, ownership, structure, privacy, and rights review; converted all 154 frozen families into transient executable fixtures; and ran focused disposable Go compatibility evidence. That review found multiple independent behavior corrections in every component and no viable retained owner graph for the stack. APG36 therefore rejects this proposal.
go-test-profile owns native Go testing and go test semantics: test
package placement, tests and subtests, helper attribution, fatal and nonfatal
failure, cleanup ordering, temporary directories and environment and
working-directory isolation, TestMain, parallelism, goroutine lifetime and
reporting, examples, benchmarks, fuzzing, selection patterns, and caching.matryer-is-test-profile owns matryer/is-specific assertion semantics after
the target repository has already selected the dependency: instance
construction and strict or relaxed mode, the available assertion families,
equality and typed-nil behavior, failure formatting, source-comment
diagnostics, and the library’s own attribution mechanism.go-cmp-test-profile owns google/go-cmp-specific semantic comparison after
the target repository has already selected the dependency: equality versus
diff, option composition and filtering, custom comparers and transformers,
ignoring and unexported-field boundaries, approximate and error comparison,
ambiguity and panic conditions, and deterministic ordering.go-testing-stack owns thin composition among the selected components: which
already-selected owner holds which part of one test’s observable contract.go-language-profile retains general Go semantics. Test-specific profiles do
not restate its contract.provisional.The first four alternatives below record APG35’s proposal-era analysis. APG36’s terminal disposition follows each premise that independent review overturned.
Native Go test profile only. APG35 rejected it because it leaves assertion and comparison behavior unowned in repositories that have already selected those dependencies, where the observed failure modes — misleading attribution, hidden semantics behind terse equality, ignore sets that quietly stop protecting behavior — are concrete and consequential.
One monolithic Go testing skill. APG35 rejected it because it binds universal native semantics to two optional dependencies, triggers on repositories that use neither, cannot be partially removed, and would carry three unrelated calibration and refresh cycles in one leaf.
Component profiles with no stack. APG35 rejected this closest alternative on the premise that a recurring cross-component question deserved its own owner. APG36 did not validate that premise: ordinary native-plus-one-component roles were already disjoint, and the narrower residual conflict cases did not justify the authored stack after the component dispositions.
Native profile plus generic assertion guidance. APG35 rejected it because generic guidance cannot state the observed library-specific facts that matter, such as a four-family assertion surface with no option surface, or an attribution mechanism that reads the assertion’s own source line rather than using the native helper mechanism.
Four owners with thin optional composition. Proposed by APG35 and rejected by APG36. Native lifecycle, matryer/is, and go-cmp source contracts each need more than the permitted one behavior correction. The stack cannot exist without the native owner and does not demonstrate independent current value.
Existing project and process owners. This is the APG36 result. It adds no new Go-testing owner while replacement component contracts and any narrower composition value remain unevidenced.
go-language-profile general Go semantics
go-test-profile native testing lifecycle and structure (always applicable)
matryer-is-test-profile optional, after project dependency selection
go-cmp-test-profile optional, after project dependency selection
go-testing-stack thin composition, only when >= 2 owners are selected
Indentation showed the intended subject-matter narrowing rather than a mandatory chain. APG36 did not adopt this graph. Existing language, repository, and process owners remain unchanged.
APG35 proposed the following three overlap controls:
APG36 found that the component facts and stack controls did not satisfy these premises. The candidate leaves were removed, and no route was added. Existing language, repository, and process owners continue to govern current work.
The proposal would have carried four refresh cycles. Native testing would
refresh on Go release
and testing API changes. matryer/is refreshes on its release, assertion
surface, equality, and attribution behavior. go-cmp refreshes on its release,
option set, documented comparer obligations, and panic conditions. The stack
would hold no independent API calibration by design and refresh only when a
boundary between owners moves.
The two optional components carried a specific staleness risk: matryer/is is at a 2023 release and go-cmp at a 2025 release, so an unnoticed upstream change would not surface through routine APG activity. Both candidate leaves treated a differing selected release as a refresh condition before any behavior claim, rather than assuming the calibrated behavior holds.
APG35 proposed that each leaf would be independently removable. The proposed mechanical removal would delete its leaf, flat projection, catalog and general-map entries, current-development release policy, strict test inventory, focused contract, and public scenario fixture while preserving evaluation and exit history. APG36 found the claimed independent usefulness false for the stack without its native owner and did not integrate any of these surfaces.
APG36 preserves the exact APG35 authoring commit in mainline history and removes the four Go candidate leaves through its forward integration commit. No projection, catalog, router, release-policy, inventory, or current fixture surface is added.
These questions require fresh maintainer authority and a new decision. They do not leave this proposal unresolved.
This ADR is Rejected. The five candidate leaves authored in APG35 are not integrated, adopted, mature, published, or deployed; APG36 removes them from the current tree after independent review and executable evidence. Integrated development remains 25 canonical skills, 25 catalog rows, and 25 flat projections. Public and active v0.3.0 remain 19/19/19.