fix(agents): bound grep execution and spill oversized tool output

[lefoulkrod]

Large tool results could overwhelm model context, and expensive Python regex searches could block the app. Oversized results now pass through ToolResultCapHook, which saves their full text to a temporary file and publishes a bounded preview and path before persistence.

The existing grep tool keeps its name, parameters, defaults, and Python matching semantics. Its async wrapper runs the matcher in a disposable process, kills and reaps it on timeout or cancellation, and retains partial matches. Search limits are explicit: 10 seconds, 1 MiB collected output, 8 MiB per file, 4,096-character line excerpts, and at most 10,000 matches. Incomplete results include retry guidance; context accepts 0–100 lines. There are no new search tools, dependency-management changes, or migration.

Validation:

• 2,027 unit and integration tests passed, including real Python worker timeout/cancellation and FakeProvider root/delegated execution with persisted temporary-file pointers.
• All 23 focused grep unit tests passed, including 30-line context compatibility.
• 5 container E2E tests passed against the rebuilt application image: oversized grep/read/shell output, reload and follow-up inspection, stuck-regex recovery, and compaction/resume. FakeProvider drives real tool execution.
• just check and just tool-docs passed.

Fixes #272.