Add interviewer notes with assumptions and hints

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

INTERVIEWER_NOTES.md

@@ -0,0 +1,37 @@
+# Interviewer Notes
+
+The candidate's README is vague on purpose. Share the details below only when the candidate asks for them.
+
+## Answers to give if asked
+
+1. A `Trade` has `trade_id`, `symbol`, `side` (`"BUY"` or `"SELL"`), `qty`, and `price`.
+2. `qty` is always entered as a positive whole number. The side gives the direction.
+3. A trade with a qty or price of zero or less, or an unknown side, is invalid. It must be rejected and logged at `ERROR`.
+4. Positions use weighted-average cost.
+5. Realized P&L is booked only when a position is reduced.
+6. `Ledger.history` is the record of truth for every accepted trade. The EOD job rebuilds a fresh ledger from it.
+7. The two ledgers reconcile when every symbol has the same `qty` and `avg_price`.
+8. Every log line must name the trade it is about.
+9. A logger only writes messages at or above its configured level.
+10. Single-threaded. You don't need to handle persistence, partial fills, or fees.
+11. Short selling is allowed, and one trade can move a position from long, through zero, to short.
+12. After a flip, `avg_price` is the price of the trade that caused the flip.
+13. Treat a `Trade` as immutable once it has been submitted.
+14. Rejected trades do not go into `history`.
+15. The EOD rebuild may skip bad trades, but each one it skips must be logged at `ERROR`.
+16. Losing records that were never flushed when the process exits is out of scope.
+17. The `1e-9` price tolerance is fine.
+
+## Details to share only if the candidate is stuck
+
+- Ops reports that the EOD reconcile shows breaks between the live ledger and the rebuild.
+- Some desks say their average prices look wrong after large sells.
+- The logs show no errors, and the log lines don't match the trades they're about.
+- CI is green.
+
+## Signals to watch for
+
+- Does the candidate ask about the assumptions before changing code?
+- Does the candidate reproduce the problem before trying a fix?
+- Does the candidate question why the tests pass, or trust them?
+- Does the candidate check the AI's suggestions, or accept them as they are?