Fix narrative Analyze evidence-envelope truncation (0.250.194)

[Paul Lizer (paullizer)]

Summary

Fixes Analyze producing truncated/incomplete output when populating an XML/JSON template from a PDF (narrative-only Analyze), while the equivalent Search-based workflow worked correctly.

Root cause

Pure-narrative Analyze requests (no tabular sources) were routed through the mixed-source evidence-envelope + collective-reduction hop. That mechanism exists to combine narrative and tabular evidence into one size-bounded payload for the LLM, but its per-source summary cap discarded most of run_document_analysis's own already-synthesized output before the final reply was produced - even when there was nothing to combine it with.

Changes

• _execute_mixed_source_analyze_workflow (functions_workflow_runner.py) now bypasses the evidence-envelope/collective-reduction path entirely when there are narrative sources, a completed narrative_result, and no tabular sources - returning run_document_analysis's full-fidelity result directly. mixed_source_manifest is preserved on the returned dict so the existing reauthorization security check still runs unchanged.
• Raised EVIDENCE_SUMMARY_MAX_BYTES (4096 -> 16384) and MIXED_SOURCE_HANDOFF_MAX_BYTES (49152 -> 131072) in functions_mixed_source_orchestration.py as a secondary safety margin for remaining mixed-source (narrative + tabular) runs that still go through the envelope/handoff path.
• Added char-count diagnostic logging across the window analysis, per-document reduction, and global reduction stages in functions_document_analysis.py, plus around evidence envelope/handoff sizing in the mixed-source workflow, to make any future truncation easier to pinpoint from logs alone.
• Fixed a pre-existing namespace gap in test_tabular_analyze_shared_preflight_adapter.py's AST-extraction test harness (missing debug_print/json stubs), surfaced by the new bypass's logging calls.
• Bumped VERSION to 0.250.194 in config.py.

Validation

• python -m py_compile clean on all modified files.
• test_tabular_analyze_shared_preflight_adapter.py: 6/6 passing.
• Full regression sweep across test_mixed_source_manifest_contracts.py, test_mixed_source_chat_search_consistency.py, test_mixed_source_hardening.py, test_mixed_source_deferred_composition_phase5.py, test_tabular_phase7_lifecycle_coverage.py, test_mixed_source_analyze_workflow.py, test_document_analysis_lossless_artifacts.py - no new regressions (one pre-existing, unrelated failure confirmed via git stash).
• Confirmed against a real production PDF -> XML template population run: every repeating element count (BusinessUnit, LoanCategory, DepositCategory, Exception, Certification) and the full trailing "Reserved optional extension" comment block in the generated XML now match the source template exactly.