Analyze delegated coverage once per test run

[wenytang-ms]

Why

Coverage for a delegated test run was collected and published from inside the loop that runs each batch of tests, one batch per project and per set of classes. A run split over several batches therefore published a report per batch. Because file coverage is stored rather than merged, the last batch hid everything the batches before it had covered.

The pre-run cleanup lived in the same loop, which made it worse in two ways: an explicit appendResult: false threw away the data the run itself had just produced, and because the directory was fixed per project rather than per run, two runs started close together could erase each other.

Analyzing no execution data at all reported every line as uncovered — indistinguishable from a run that genuinely covered nothing. A run whose agent never attached looked like a real 0%.

What

One directory per run, analyzed once. A delegated run gets a directory of its own, and every project's data is analyzed after all batches have finished rather than after each one.

appendResult becomes a matter of scope, not deletion. Analyze only this run's directory, or the root that still holds the runs before it. Nothing is deleted while a run might be writing into it, which is what let concurrent runs erase each other. Old run directories are collected on a time bound instead, and a directory younger than that bound is never touched.

No execution data is an error. A run that produced nothing reports the failure rather than a fabricated 0%.

The contract

Delegated coverage now goes through an explicit request rather than a bare output path:

interface CoverageRequest {
    format: 'jacoco-exec';
    version: 1;
    executionDataDirectory: vscode.Uri;
    agentJar: vscode.Uri;
}

• The agent jar comes from here, so the agent that writes the data and the analyzer that reads it are always the same version. A runner does not choose, resolve or configure one.
• This extension owns the directory — it creates it, decides how widely to analyze, and cleans up. A runner writes .exec files into it and must not clear it or write with append=false.
• Nothing about reporting is in the payload: no XML, no source directories, no report tasks. The runner produces execution data; analysis happens here.
• Cancellation is IRunTestContext.cancellationToken, not part of the coverage request.

Runners advertise support with the jacoco-exec@1 capability, so the version is negotiated rather than assumed.

Testing

• tsc and eslint clean.
• mvnw clean verify builds the Java plugin clean, including checkstyle.
• New CoverageHandlerTest case covers the no-execution-data failure.