Add market data backend code review

[dobidobi77-bot]

Comprehensive review of backend/app/market/ and its test suite, written to planning/MARKET_DATA_REVIEW.md. Supersedes the round-1 review in planning/archive/.

Test results

All round-1 action items were applied and the suite is green: 73 passed, 91% coverage, ruff check clean.

Two defects the passing suite does not catch

Both verified empirically, not inferred from reading.

1. The Massive (real market data) path is 100% non-functional. massive_client.py:103 reads snap.last_trade.timestamp — a field absent from the SDK's LastTrade model (it's sip_timestamp). The AttributeError is swallowed by the skip-malformed-snapshot handler two lines down, so every well-formed snapshot is dropped and the cache is never written. Set MASSIVE_API_KEY and the whole app has no prices. The same lines also divide by 1000 assuming milliseconds; Polygon timestamps are nanoseconds.

It survived testing because test_massive.py's _make_snapshot builds a MagicMock and assigns the invented field — test and code share one wrong assumption. Coverage on that module is 94%.

2. Shock events swamp the GBM engine by ~317x in variance. With event_probability=0 the math is exact (AAPL/GOOGL correlation 0.598 vs target 0.60; realized sigma 0.221 vs configured 0.22). At the shipped 0.001, correlation collapses to 0.003 and realized volatility runs 8-28x configured and roughly equal across all tickers — so the per-ticker sigma tuning and the entire Cholesky correlation design are inert. Prices wander 9.2%/hour instead of 0.7%.

Also covered

• Verified SSE wire format documented for the frontend, closing PLAN A5 — unnamed events, one frame keyed by ticker, change_percent included
• No SSE keepalive: 6s idle yields 520 bytes, all from the initial burst
• Ticker normalization diverges between the two MarketDataSource implementations
• Module-level router singleton defeats the cache injection its own docstring promises
• Findings against PLAN A1/A2/A3, including that TickerSnapshot.prev_day and todays_change_percent would close A1 for free
• A section on why 73 tests and 91% coverage missed both defects, and what to change structurally

Documentation only — no source changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VGUSZLyF2LLqRMz18yg2Re