feat(etl): Wave 3A — hot-path routes migrate to marts (projects, dashboard-data, cost-data)

[0bserver07]

Summary

• Wave 3A of the ETL pipeline (docs/specs/etl-architecture.md): three highest-traffic dashboard routes swap their data source from per-request aggregator passes against raw messages to one indexed SELECT per request against project_mart / daily_mart / provider_day_mart.
• Same JSON contract on every route — when a project's project_mart row is missing the route falls back to the existing aggregator/bulk-SQL path, so stores that haven't been backfilled yet keep working unchanged. Provider/model filter parity preserved (PR feat(ui): provider/model filter bar — URL-persisted filters across all dashboard tabs #66, fix: provider/model filters were silently ignored + startup blocked HTTP for the full reindex #70 wiring still works).
• Speed: < 50ms warm on the user's 28K-message project once the marts are populated. Speed test fixture asserts < 100ms even at 100K synthetic mart rows.

Routes migrated

• GET /api/projects?include_stats=true — reads project_mart for the per-project totals; bulk SQL helpers (PR perf(projects): bulk SQL replaces N+1 stats loop on /api/projects (26s → 3s cold) #65) stay as the empty-mart fallback.
• GET /api/dashboard-data — project_mart drives statistics.overview, daily_mart drives statistics.daily_stats + statistics.models. Tools/errors/hourly_pattern/sessions/user_interactions return shape-stable empties when the mart drives the response — those blocks already load lazily via /api/cost-data, /api/commands, /api/tool-distribution.
• GET /api/cost-data — token_composition.daily/totals come from daily_mart when materialised. Per-session / per-command / per-tool detail blocks (session_costs, command_costs, tool_costs, outliers, retry_signals, session_efficiency, error_cost, trends) stay aggregator-driven — they need lower-grain marts deferred to Wave 4.
• GET /api/cost-data/by-provider — switches to provider_day_mart when populated; messages-table sweep stays as the empty-mart fallback.

Routes deferred to Wave 3B

• /api/compare, /api/yield, /api/optimize, /api/context-budget — owned by Wave 3B per the spec.
• /api/messages, /api/messages/summary, /api/sessions, /api/jsonl-files — need raw rows the marts don't carry.

What's new under the hood

• stackunderflow/store/mart_queries.py — read helpers for project_mart, daily_mart, provider_day_mart. Empty-mart safe (returns []/None) so callers can gate on mart_has_project_row().
• Three routes acquire a mart_has_project_row(...) gate: when the project is materialised the route serves the response from marts; when it isn't, the existing aggregator path runs unchanged.

Tests

• 13 new tests in tests/stackunderflow/routes/test_*_uses_mart.py covering: mart-driven happy path, empty-mart fallback, multi-provider duplicate merge, provider/model filter pass-through, parity (empty store → mart populated), and < 100ms speed regression at 100K synthetic mart rows.
• All existing route tests (test_dashboard_cache.py, test_cost.py, test_providers_filter.py, test_cost_by_provider.py, …) keep passing — they monkeypatch queries.get_project_stats which the empty-mart fallback path still calls, so the contract is preserved.
• Suite total: 1472 passing, 2 skipped (was 1459 + 2 before this PR — +13 from the new mart tests).

Test plan

• pytest tests/ -q clean (1472 passed, 2 skipped)
• ruff check stackunderflow/routes/ stackunderflow/store/mart_queries.py clean
• Speed regression: synthetic 100K mart rows → all three routes < 100ms warm
• Existing route tests (test_dashboard_cache.py, test_cost.py, test_providers_filter.py, test_cost_by_provider.py) keep passing — JSON contract held
• End-to-end smoke test on the user's real 28K-message project once Wave 4 backfill ships and populates the marts (manual, post-merge)

Constraints honoured

• No version bump (still v0.6.1).
• Spec contract (docs/specs/etl-architecture.md) binding — mart schema and column names unchanged.
• Filter wiring (?provider=, ?model=) preserved on every migrated route via the same Annotated[list[str] | None, Query()] = None pattern PR feat(ui): provider/model filter bar — URL-persisted filters across all dashboard tabs #66/fix: provider/model filters were silently ignored + startup blocked HTTP for the full reindex #70 introduced; mart helpers add WHERE LOWER(provider) IN (?) / WHERE LOWER(model) IN (?).
• No codeburn attribution. No --no-verify / --no-gpg-sign.