fix(windows): avoid ENAMETOOLONG in Claude reviews

[rei-okawara-portx]

Background

cc:review can inline up to 64 KiB of working-tree context. The Claude CLI wrapper previously placed that entire prompt in spawn() argv, which exceeds the Windows process command-line limit before Claude starts.

Problem

• Large reviews fail with spawn ENAMETOOLONG on Windows.
• Calling the npm claude shim directly is unreliable from Node on Windows because PATH normally resolves to a .cmd wrapper.
• The previous native executable lookup assumed the default home-based npm prefix and missed relocated APPDATA, custom npm prefixes, and PATH-selected installations.

Solution

• Keep CLI options in argv and send the prompt through Claude Code's documented non-interactive stdin input.
• Record stdin pipe failures and fail the job closed instead of reporting a review without a delivered prompt.
• Resolve the native Windows claude.exe in this order:
  1. CC_PLUGIN_CODEX_CLAUDE_BIN
  2. native claude.exe shims directly present in PATH
  3. npm prefixes present in PATH
  4. npm_config_prefix
  5. APPDATA\\npm
  6. the standard home npm prefix
  7. normal claude command lookup
• Deduplicate Windows search roots case-insensitively.
• Add unit, integration, and E2E coverage for stdin prompt delivery and binary discovery.

Verification

• npm run test:cross-platform: 260 passed, 0 failed
• node --test tests/claude-cli.test.mjs: 92 passed, 0 failed
• npm run lint: passed
• npm run typecheck: passed
• npm run check:version-sync: passed
• npm run check:changelog: passed
• git diff --check: passed
• Claude Code 2.1.170 completed a real 41,922-byte working-tree review with job status completed, phase done, and no job error.
• A final installed-cache smoke task returned exactly OK with exit code 0.
• Standard and adversarial Claude reviews were run. The adversarial review identified an unnecessary --input-format text compatibility risk; that flag was removed because text is already the documented default.

Scope / out of scope

• No shell invocation was introduced.
• No prompt temp files or environment-variable transport were introduced.
• Review sandbox/tool policy and tracked-job behavior are unchanged.
• Migrating the wrapper to the Agent SDK is intentionally out of scope; it would require redesigning auth, session, sandbox, and job control for no benefit to the one-shot prompt transport.

Remaining risks

• The local full npm test command still includes existing platform-specific tests that assume POSIX bash, /proc, or executable test shims and is not green on native Windows. The repository's cross-platform suite is green.
• Local Claude Code reported that srt-win.exe was unavailable and disabled its OS sandbox. Standard review still exposed no Bash, Write, or Edit tools; fixing the sandbox runtime distribution is a separate issue from prompt transport.