feat(tools): audit, fix three bugs, add multi_edit / glob_files / todo_write

[Pterjudin]

Summary

Outcome of a full read-through audit of the agent tool surface. Fixes three verified bugs and adds three missing tools that close gaps versus Claude Code / Cursor / OpenCode.

Bug fixes

• search_pathnames_only silently dropped `include_pattern` — validator destructured `params.search_in_folder` (which the model never sends) while the LLM-facing parameter name is `include_pattern`. Every include_pattern argument was discarded, returning unfiltered results.
• terminalToolService silent degradation on missing CommandDetection — when shell-integration capability failed to mount within 10s, the resolve path was skipped, leaving `waitUntilDone` as a dead promise. Commands appeared to complete but returned stale buffer contents. Now throws a clear error instead.
• Removed dead pathname_search comment block in prompts.ts — easy to mistake for a live definition while reading the file.

New tools

• `multi_edit` — atomic multi-block search/replace on a single file. Pre-checks every `old_string` exists before applying anything; if any miss, no edits are written. Supports per-edit `replace_all`. Approval class: `edits`. Builds on the existing applySearchReplaceBlocks engine.
• `glob_files` — file listing by glob pattern, sorted by modification time newest-first. Returns mtime + size per file. Limit clamped 1–1000. Closes the gap with Claude Code's Glob — `search_pathnames_only` is substring-only and unsorted.
• `todo_write` — model-managed task list, per-session. Enforces a single `in_progress` task at a time. Stored array exposed via `ToolsService.getLatestTodos()` for future UI rendering.

Audit deliverable

Companion audit document at `TOOLS_AUDIT_2026-05-25.md` in repo root (not committed in this PR — intended as a working reference, not project doc). It includes the full 29-tool inventory, competitor matrix, design issues with the four "actor" tools (`extract_function` / `rename_symbol` / `automated_code_review` / `generate_tests` describe taking actions but only return data), and recommended follow-ups.

Verification

• `npm run compile-check-ts-native` — clean
• `npm run valid-layers-check` — no new violations (pre-existing IMainProcessService warnings are unchanged)
• Manual trace: each new tool has a type entry, validator, impl, stringOfResult, prompt description, and approval-class entry where applicable

Follow-ups (not in this PR)

• Unit tests for validators — requires extracting validators out of the `ToolsService` class so they can be exercised without full DI. Refactor preferred as separate PR.
• Honest descriptions for `extract_function` / `rename_symbol` / `automated_code_review` / `generate_tests` — these tools describe actions they don't perform. Either rewrite descriptions or implement the actions. Discussion needed.
• SSRF guard on `browse_url` — no private-IP allowlist; covered by the separate security review.
• `run_nl_command` YOLO substring heuristic — naive and trivially bypassable; recommend removing until a real classifier exists.
• UI surface for `todo_write` — sidebar widget consuming `getLatestTodos()`.