Stop opening a null-device handle per Capture and spinner task

[GoodForOneFare]

Bug

Capture#initialize and SpinGroup#add both defaulted duplicate_output_to: to a freshly opened File::NULL handle:

duplicate_output_to: File.open(File::NULL, "w")  # Capture
duplicate_output_to: File.new(File::NULL, "w")   # SpinGroup#add

Every capture therefore opened an OS handle solely to discard writes, including one for every spinner task. Ruby closes these handles when their File objects are finalized, but they remain open while captures or queued task closures are live.

Verified on main: constructing 10 captures opens 10 new descriptors.

Fix

Default duplicate_output_to: to nil and skip the duplicate write when unset. This avoids opening a descriptor just to discard output.

The option is typed as io_like? (IO | StringIO | nil) so its signature matches the streams accepted elsewhere in cli-ui. Explicit duplicate streams continue to receive captured output and remain caller-owned. An IOError from a closed duplicate stream is ignored, matching the module-level duplicate writer.

The module-level StdoutRouter.duplicate_output_to used by CLI::UI.log_output_to is a separate mechanism and is otherwise unchanged.

Tests

• Ten simultaneously live Capture objects add no descriptors.
• Twenty pending spinner tasks add no descriptors while queued or after completion.
• Explicit StringIO duplicates receive captured stdout through both Capture and SpinGroup#add and remain caller-owned.
• A merged-output capture duplicates stderr.
• A duplicate stream closed during capture does not interrupt the caller block.

The descriptor tests inspect /dev/fd on macOS and Linux and skip where it is unavailable. They compare descriptor sets and assert only that no descriptors were added, so unrelated File objects finalized between samples cannot make the test fail merely because the total count decreased.

🤖 Generated with Claude Code

(posted by an LLM bot on behalf of Gord)