fix(agents): handle find subprocess stream failures - #101158

Related: #101020

What Problem This Solves

Fixes an issue where agent runs using the find tool could crash and leave fd running when its stdout or stderr stream failed.

The contributor PR carrying this fix was automatically closed only because its author exceeded the repository's open-PR limit. This maintainer-owned replacement preserves the contributor's authorship and reviewed implementation.

Why This Change Was Made

The fd subprocess now has one idempotent stream-failure path shared by readline, stdout, and stderr. That path stops the child, closes readline, removes abort handling through the existing settle owner, and rejects exactly once even though Node forwards an input error through both readline and the raw stream.

User Impact

The find tool now fails the individual tool call cleanly on an OS-level pipe/read failure instead of risking an agent-process crash or orphaned search process. Successful searches, cancellation, result limits, and existing fd exit handling are unchanged.

Evidence

• Node behavior probe confirmed an input-stream failure is emitted first by readline and then by the raw stream; the settled guard keeps cleanup one-shot.
• Focused type-aware oxlint: passed.
• Focused oxfmt --check: passed.
• git diff --check: passed.
• Fresh high-reasoning autoreview: no actionable findings (correctness confidence 0.87).
• Exact-head CI run 28818145434: passed.
• Native OPENCLAW_TESTBOX=1 scripts/pr prepare-run 101158: passed at cf5000f1de7042a9a0aa8cc7a9344b840aeaa0f6.

AI-assisted maintainer repair. I reviewed the full find subprocess lifecycle, its registration callers, the sibling grep implementation, adjacent tests, and Node's readline error contract.