feat(etl): Wave 4A — compare/yield/optimize/messages-summary migrate to marts (3B redo)

[0bserver07]

Summary

Wave 4A finishes what Wave 3B started: migrates four analytical surfaces from the per-request aggregator pass over raw messages onto the ETL mart layer shipped in Wave 1 + 2 + 3A. Empty-mart fallback to the aggregator path is preserved on every migrated surface so users on un-materialised stores keep working.

• services/compare.compare_models() reads model_day_mart for per-(model) totals (calls, tokens, cost) and session_mart for primary-model attribution (sessions, one-shot %, retry rate, $/session). When project_filter is set the call drops to the aggregator path because model_day_mart does not carry project_id.
• services/yield_tracker._query_sessions() reads session_mart for the session list (cwd / started_at / cost_usd / primary_model). Cwd is still pulled from messages.raw_json since the v1 session_mart leaves cwd NULL per the builder docstring.
• reports/optimize._detect_cache_overhead() reads pre-summed per-session input_tokens + cache_create from session_mart. The other six detectors stay on the aggregator path — docstrings now spell out why (signals like tool_use blocks, raw_json parsing, per-message byte counts aren't on any mart yet).
• routes/data /api/messages/summary's top-level total (and a new bonus total_sessions field) come from project_mart. The detail blocks (by_type / by_model / total_tokens) still use the messages list because those dimensions aren't materialised.
• routes/yield_route.get_yield gains the same Query sentinel sanitization routes/compare.py already had so direct test calls don't leak the FastAPI sentinel into the service layer.

store/mart_queries.py grows new helpers (kept consolidated, no inline SQL in routes): model_day_totals, session_mart_rows_for_compare, session_mart_rows_for_yield, session_mart_cache_overhead, project_mart_messages_summary_totals, plus mart_has_session_rows / mart_has_model_day_rows existence gates.

Test plan

• pytest tests/ -q — 1489 passed, 2 skipped (was 1472 pre-PR; +17 new)
• ruff check stackunderflow/services/compare.py stackunderflow/services/yield_tracker.py stackunderflow/reports/optimize.py stackunderflow/routes/data.py stackunderflow/routes/yield_route.py stackunderflow/store/mart_queries.py — all pass
• ruff check on the 4 new test files — passes
• Compare provider attribution: codex sessions surface as provider="codex" from session_mart, not the legacy JOIN projects.
• Compare provider filter: provider_filter="claude" excludes codex entries even when model_day_mart has rows for both.
• Yield cwd parity: session_mart.cwd=NULL does not break the cwd fetch — the service still pulls cwd from messages.raw_json.
• Yield project_filter: slug-based filter pushes into the mart's JOIN projects clause.
• Optimize cache_overhead: severity ladder + finding shape (incl. details.sessions schema) identical between mart and aggregator paths.
• Messages-summary: total_messages=4242 in project_mart overrides the messages-list count of 1.
• Empty-mart fallback: each migrated surface has a parity test that seeds the legacy tables only and confirms the aggregator path still answers correctly.
• Speed: 50K mart rows answer compare / yield / cache-overhead within budget.

🤖 Generated with Claude Code