Calibrate accept confidence fallback

[yha9806]

Summary

• allow low-confidence accepted decisions to stay with tiny_model when required source context is available
• keep explicit fallback_to_agent and missing-source fallbacks unchanged
• update training-effectiveness and residual-audit expectations after the tiny-router confidence gap is removed

Verified effect

• public source-context dry-run fallback: 21 -> 6 (reduction 15)
• private recovery dry-run fallback: 6 -> 4
• recovered no_source_context_for_required_source: 2 -> 0
• residual audit tiny_router_candidate_count: 0
• residual fallback_agent_count: 4

Verification

• PYTHONPATH=src python3 -m pytest tests/test_dry_run_decision_router.py::test_dry_run_decisions_do_not_fallback_accepted_cases_with_source_context tests/test_training_effectiveness_report.py::test_training_effectiveness_report_compares_source_context_router_improvement tests/test_fallback_residual_audit.py::test_fallback_residual_audit_classifies_remaining_agent_work -q
• PYTHONPATH=src python3 scripts/real_source_context_recovery_eval.py --repo-root . --case-source-manifest docs/benchmarks/learning/combined_case_source_manifest_v1.json --artifact-search-root /Users/yhryzy/dev/vulca/.scratch/p2b-live-shipgate --image-search-root /Users/yhryzy/dev/vulca/.scratch/p2b-live-shipgate --source-dependency-manifest docs/benchmarks/learning/real_source_dependency_label_manifest_v1.json --output-dir build/tiny_confidence_calibration_v1/real_recovery_eval --report build/tiny_confidence_calibration_v1/real_recovery_eval/report.json --max-recovered-source-context-gaps 0 --min-fallback-agent-reduction 2 --min-recovered-eval-cases 2
• PYTHONPATH=src python3 scripts/training_effectiveness_report.py --repo-root . --output-dir build/tiny_confidence_calibration_v1/training_effectiveness --report build/tiny_confidence_calibration_v1/training_effectiveness/report.json
• PYTHONPATH=src python3 scripts/fallback_residual_audit.py --decision-path build/tiny_confidence_calibration_v1/real_recovery_eval/recovered/dry_run/dry_run_decisions.jsonl --report build/tiny_confidence_calibration_v1/residual_audit/report.json
• PYTHONPATH=src python3 -m pytest tests/test_dry_run_decision_router.py tests/test_training_effectiveness_report.py tests/test_real_source_context_recovery_eval.py tests/test_fallback_residual_audit.py tests/test_source_context_gap_pack.py -q
• ruff check src/vulca/learning/dry_run_decision_router.py tests/test_dry_run_decision_router.py tests/test_training_effectiveness_report.py tests/test_fallback_residual_audit.py
• python3 -m py_compile src/vulca/learning/dry_run_decision_router.py
• git diff --check