fix(tools): thrown tool errors match original wire+display format

[ericleepi314]

Summary

A Read of a directory printed raw wire markup in the transcript:

⏺ Read(src)
  ⎿  Error: <tool_use_error>path is not a file: /Users/…/obama-blog/src</tool_use_error>

The original Claude Code shows Error: EISDIR: illegal operation on a directory, read '/…/src' — no tags. This port had missed that on two independent layers, plus two supporting gaps:

1. Wire parity — thrown call() errors are sent RAW in the original: content = formatError(error) with is_error: true (toolExecution.ts:1633+1674). <tool_use_error> tags are reserved for pre-execution failures (no-such-tool, InputValidationError, validateInput) — those keep their tags, identical to TS. _format_error is now a faithful formatError port (stderr/stdout parts, Command failed with no output fallback, 40KB middle truncation, AbortError→message-or-interrupt-constant).
2. Display parity — ported FallbackToolUseErrorMessage into formatToolResult: verbatim extractTag (utils/messages.ts:635), sandbox-violation/<error>-tag stripping, non-verbose InputValidationError → Invalid tool parameters, plus each tool's renderToolUseErrorMessage collapse (File not found, Error reading file, File must be read first, Error searching files, Error writing file, Error editing notebook) and the pluralized overflow hint.
3. Error signal on OpenAI-family wires — role=tool / function_call_output / FunctionResponse have no is_error field; our converters never consumed it, so the tags were the only failure marker those models got. All three (chat-completions, Responses, Gemini) now prefix Error: on is_error results, mirroring convertToolResultContent (openaiShim.ts:309). Load-bearing for the de-tagging.
4. Read message parity — directory reads raise the byte-exact TS message (readFileInRange.ts:89): EISDIR: illegal operation on a directory, read '<path>'. Other non-files (FIFOs/sockets) keep the generic refusal (TS streams them; deliberate pre-existing divergence).

Verification

• pytest: 8529 passed / 0 failed (15 new in tests/test_tool_error_format_parity.py)
• vitest: +8 passed; failures byte-identical to main's pre-existing baseline; tsc --noEmit clean
• Critic review: APPROVE — every port verified against the TS source; adjacent converter+loop-guard suites re-run (170/170); independent JS harness of the ported display logic 25/25

Non-gating follow-up notes (from review)

• tool_execution.py:520 abort-path content pre-bakes Error: → double prefix on OpenAI-family wires for resumed-session aborts (cosmetic; the clean fix is aligning that content with formatError, out of scope here)
• ctrl+o expanded view shows model-facing raw bytes for the still-tagged class (existing TUI convention vs TS verbose stripping); stripUnderlineAnsi not ported

🤖 Generated with Claude Code