fix: tolerate benign trailing prose after final tool request

[sendralt]

Problem

When an agent emits a fully valid tool-request JSON followed by stray non-JSON text, the strict equality gate in extract_tool_request (root != content) rejects the whole message as misformat, the tool never executes, and the run degrades into repair loops.

A naive relaxation (accepting any prefix) is unsafe: streaming early-stop and repair routing rely on strict rejection of trailing text, e.g. tests/test_tool_request_normalization.py asserts trailing-text rejection and stream snapshots must not stop early while a second tool call may still be streaming.

Solution

Boundary-scoped leniency:

• NEW helpers/extract_tools.py extract_tool_request_final() - accepts a complete tool request followed by benign non-JSON trailing prose, while still rejecting preamble text before the request, concatenated additional JSON roots, and trailing text containing braces, brackets, or double quotes so those cases keep routing to repair instead of executing partially.
• agent.py process_tools now calls the final variant for completed messages. Streaming snapshots keep the strict extract_tool_request boundary, so early-stop semantics are untouched.

Testing

• New regression test test_extract_tool_request_final_tolerates_trailing_prose covering: trailing prose accepted; preamble rejected; concatenated second root rejected; brace/bracket-bearing suffix rejected; truncation rejected; plain JSON and prose rejected.
• All existing assertions in test_tool_request_normalization.py and test_stream_tool_early_stop.py remain green (strict behavior unchanged).
• Verified zero regressions via baseline comparison of failing-test sets before and after the change (identical sets).