ci: enforce a shared strict repository quality gate

This commit is contained in:
2026-10-14 12:14:11 +01:00
parent 461b554c1a
commit 418e574ac8
14 changed files with 454 additions and 118 deletions
@@ -142,3 +142,40 @@ must not be presented as completed strict workspace qualification.
- Coverage gates that live in repository helpers are easier to evolve, audit, and keep consistent than CI-local logic.
- Audit source ownership and evidence validity before comparing percentages.
A passing percentage is useful only when its denominator has a clear owner.
## A Shared Gate Must Share Failure Semantics Too
The strict-gate correction derives from DEC-0063. A common coverage helper did
not establish common CI policy while one pipeline only ran permissive lint/tests
and another produced domain reports. Both CI consumers now provision the declared
tools and invoke the same Makefile-owned `ci-domains` target. Strict Clippy covers
workspace targets and features with warnings denied; concrete lint fixes replace
warning suppression. The existing coverage map and floors remain unchanged.
The Makefile owns the ordered steps. A serial runner performs format, strict lint,
all-target tests, the five named debugger socket cases, fresh coverage/export and
all domain audits/gates. Serial child Make invocations avoid races between clean,
Cargo execution and report generation. A second independent gate in the same
checkout still requires operator coordination.
Failure publication is part of the contract. Before a run, old reports are
removed; every executed step receives a log, exit status, revision and source/config
identity. A later failure retains current partial reports. An early failure cannot
archive old HTML as current success. Both CI providers publish available evidence
on failure and preserve the gate's nonzero status. Regression fixtures exercise
failed execution, success and failure after partial report generation.
Ignored tests are not one qualification category. Five localhost socket tests
belong in automation, but physical audio, endurance, packaged installation and
manual timing/layout probes need their own named execution. Restricting the socket
filter prevents accidentally turning a device requirement into a CI prerequisite.
The local macOS/aarch64 strict gate passed 1,039 ordinary tests, five socket tests,
three runner regressions and all fifteen coverage gates; fresh coverage separately
replayed 1,039 tests. The ordinary phases reported thirteen ignored tests; five
of these ran in the separate socket step, while eight remained unselected. The initial sandbox
socket refusal is distinguished from the successful authorized run. Pipeline
configuration was inspected, but no remote or Linux gate was executed. Platform
equivalence concerns required commands, not identical measured coverage. See the
[published operational contract and evidence](../../../docs/vm-arch/COVERAGE_DOMAINS.md)
for prerequisites, manual commands and practical limits.