ServiceContenderSpawnerTests: measure whether the keyless [NotInParallel] is still needed #330

Owner: Platform lane.

What

ServiceContenderSpawnerTests (tests/OpenCode.Sdk.Tests/Hosting/BackgroundService/Contender/) carries the keyless [NotInParallel], so the class runs alone after every other test. The comment that justified it was false and has been corrected in #327. Nobody has measured whether the class still needs the key.

Why the old rationale no longer held

The comment said the release-drain proof "paces 1024 timer waits that lag under module overlap" on the three-vCPU macOS leg. Since f6821b0 ("the release-drain proof holds its writer on a file the test creates, instead of pacing it on the clock"), the probe writes 4 KiB chunks and waits on a hold file only after the first chunk. Its only timer loop is the hold-file poll, and the poll's interval decides only how soon the writer notices, never whether a proof passes. The comment also said "every test starts a real process", which the missing-executable test does not.

This was found by the cross-family review of #327 (Sol 1M high, dispatch run 2026-10-15-layout-f0-residue-review, finding 1). #327 now states only what holds, and leaves the key in place.

Acceptance

A measured answer, not an argument:

• Run the class with the keyless [NotInParallel] and without it (with the server-process key, or none) under the gate's load, on Windows and on the three-vCPU macOS leg, enough runs to see a timing failure if one exists.
• Keep the key only if the measurement shows the class needs it, and write the measured reason into the class comment.
• If it goes, the class runs in parallel with the rest of the suite, and the main pass's wall time records the gain.

Per the repository's rule, a flaky outcome is fixed at its root cause, never by a retry or a wider bound.