[Kapozux]

Model validation drops numeric zero and boolean false from additional parameters, so the configuration checked can differ from the configuration used at runtime. Replace the truthiness filter with explicit missing/empty-string checks, preserving zero, false and supplied empty containers while continuing to skip empty keys, None and empty strings.

Fixes #1446.

Validation:

• Python 3.11.11: 18 new offline regression cases execute the production handler, real request DTOs, parameter parser, LLMConfig and StreamingResponse with a substituted model transport. On the unmodified base, 6 fail and 12 pass; with the fix, all 18 pass.
• New regressions plus existing supplier configuration checks: 40 passed.
• Ruff check and format check pass for the new test file; git diff --check passes. The production file retains the same 18 existing Ruff findings and requires formatting both before and after this change; no unrelated cleanup is included.

This concerns typed API/configuration values; manually typing "0" into the current text input produces a string and does not reproduce the zero-value case. No live authenticated HTTP route, database, browser or external LLM was exercised. The existing LLMConfig Pydantic deprecation warnings remain.