0109 — The cancellation token was never an overload, and 0107 said it was

Status: accepted · Date: 2026-09-11

Context

Decision 0107 moved the suites to xunit v3 and suppressed xUnit1051 — pass TestContext.Current.CancellationToken to a call that accepts one — for the whole tests/ area. It gave a reason, in its evidence section and again in a consequence:

Eight of those are PersistenceOverloadTests, whose whole subject is that the overload without a token exists and behaves like the one with it, so a blanket fix deletes what they assert.

That is false, and it was asserted without reading the file. PersistenceOverloadTests says what it is about in its own summary:

Every vectorizer ships four ways to persist — stream, file, and an async counterpart of each — and the round-trip suite exercised only the stream ones.

The overloads under test are stream against file and synchronous against asynchronous. The eight sites are SaveAsync(stream) and LoadAsync(stream) calls, and the token is not an overload at all: measured across src/, 20 methods take CancellationToken cancellationToken = default and zero take it as a separate mandatory overload. Passing it calls the same method by the same path and asserts exactly the same thing.

So the suppression had no surviving justification, and the follow-up #647 was scoped around a fiction: it was written to fix 46 of 54 sites and leave 8 with a #pragma each.

Decision

All 54 sites take the token, and xUnit1051 comes out of tests/Directory.Build.props. Not one site needed a #pragma: the build is clean with the rule on, and warnings are errors here. 6942 tests across 32 assemblies before and after, and the diff is 54 insertions against 54 deletions — one line each, nothing else moved.

The argument is named rather than positional — cancellationToken: TestContext.Current.CancellationToken — because most of these methods take an options parameter before it, so the positional form would bind the token to ArtifactLoadOptions? or EncodingOptions? and fail to compile. That is worth recording as the shape a later site should copy, not rediscover.

Why it happened, which is the part worth keeping

The claim came from the file's name. PersistenceOverloadTests reads like a suite about overloads, the rule was about an overload-shaped thing, and the two were joined without opening the file — whose first twelve lines say otherwise.

Decision 0074 exists to stop exactly this, for external libraries: read the exported surface, never the README. The protocol was followed for datasketch, Microsoft.ML.TimeSeries, Accord.Statistics and a dozen others, and not applied to a file in this repository — where it is cheaper, not more expensive.

An ADR is a record of evidence. One stating an unread file's contents as measured is worse than one stating nothing, because the next reader has no way to tell the two apart. This record is the correction; 0107's decision — xunit v3 whole, coverage as Visual Studio XML — is untouched and stands.

Consequences

• xUnit1051 is enforced. A new call that accepts a token and is given none fails the build, which is where the rule is worth having.
• 0107's evidence section is wrong where it names PersistenceOverloadTests, and right everywhere else — the churn measurement, the global.json constraint, the coverage format and the two pinned-name traps were all measured. docs/decisions/index.yaml carries the amends edge so a reader of 0107 reaches this.
• The habit this repository already has for external surfaces — read it, do not infer it — applies to its own files. Nothing enforces that and nothing can; it is written here so the next claim about a test file's subject is checked against its summary.