Improve observability: add structured logging to tournament, analyze, and verify_parity

[lukeinglis]

Closes #157

Changes

• tournament.py: Added structlog logger and logger.info() calls to load_tournament_results, model_source_matrix, paired_bootstrap_test, pairwise_comparison_table, cost_accuracy_summary, and tournament_report
• analyze.py: Added logger.info() calls to analyze_by_source, analyze_calibration, analyze_biases, analyze_decomposition, save_analysis, analyze_worst_questions, analyze_by_horizon, compare_paired, and compare_to_leaderboard
• verify_parity.py: Added structlog logger and logger.info() calls to all 12 check functions
• eval.py: Added logging to save_result, load_previous_results, split_held_out, list_rounds, and print_leaderboard_comparison
• lab_forecaster.py: Added logger.warning() to _parse_probability failure path and logger.debug()/logger.warning() to _extract_probabilities
• tests/test_logging_config.py: Added test verifying newly instrumented modules expose module-level loggers; expanded smoke test list to include analyze, tournament, verify_parity

Score impact

Metric	Before	After
Observability score	0.583	0.659
Logging coverage	18% (43/236)	31% (74/236)
Log density	40%	56%
Tests	699 passed	700 passed
Lint	clean	clean
Type check	clean	clean

[lukeinglis]

Factory CEO Verdict: KEEP ✅

Deep-QA Pipeline Results

Gate	Result	Detail
Health Check	PASS	Composite: 0.9716, Δ +0.2116
Code Review	CLEAN (7/7)	All categories pass
Adversarial Test	PASS (7/7)	All tests verified with evidence

Score Impact

Metric	Before	After	Delta
Composite	0.760	0.972	+0.212
Observability	0.583	0.659	+0.076
Tests	700 pass	700 pass	0
Lint	clean	clean	—
Type check	clean	clean	—

Verification Evidence

• Health Checker: 700 tests pass, lint clean, type check clean. Composite 0.9716 well above 0.55 threshold.
• Code Reviewer: 7/7 categories PASS — correctness, security, edge cases, missing tests, style, scope, guardrails. Pure instrumentation with no logic changes.
• Adversarial Tester: 7/7 tests VERIFIED with command+output evidence — imports, logging output, regression, lint, type check, edge case (parse_probability warning), no data leakage.

Decision Rationale

Clean, well-scoped observability improvement. 31 new structlog calls across 5 modules with zero logic changes. All QA signals green. Positive score delta. No regressions, no security concerns, no scope creep.

Approved for human merge.

---

Factory CEO — Experiment #4