test(windows): read windowsHide the way an object literal evaluates

[bompus]

child-process-windows-hide.test.ts checks that every child_process spawn under src/ sets windowsHide: true. Its object-literal reader only let a spread turn the option on, so { windowsHide: true, ...{ windowsHide: false } } passed although the child would open a console window. This came up in review of #359.

The reader now applies members in source order, the way the literal evaluates:

• A spread clears an earlier true unless it provably has no windowsHide key.
• A computed key, shorthand, accessor or method that may set windowsHide clears it too. Quoted keys count.
• Cycle detection follows the bindings being expanded, so spreading one value twice is no longer taken for a cycle.
• A parameter is checked against every call. An omitted argument takes the parameter's default, and a spread argument that may supply it makes the parameter unprovable.
• A loop variable no longer borrows an outer binding's proof.

Variables are still read from their initializers, so a later mutation or a getter's side effect goes unseen. The doc comment says so: the guard catches a forgotten option, not code written to evade it.

Test-only change: no CHANGELOG entry. README rows checked: none describe this guard.

Checks

• The guard test passes, 3 tests. All current src/ spawn sites pass (36 proven, 2 allowlisted).
• The new cases fail against the previous reader, and the default, spread-argument and loop cases each fail with their fix removed.
• Astra xhigh review: PASS in round 4. Rounds 1–3 found the computed-key, shared-reference, default-argument, loop-shadowing and spread-argument cases fixed here. Round 2 also found two limits that are accepted and documented.