[hey-research-lab]

What

SignerClient.withdraw, transfer, transfer_same_master_account, update_margin and create_market_order_quote_amount convert a human amount to ticks with int(amount * scale). Binary floating point makes many ordinary amounts land just below the integer, so the truncation signs one tick less than asked: withdraw(ASSET_ID_USDC, ..., 2.01) signs 2009999 (2.009999 USDC), and withdraw(ASSET_ID_ETH, ..., 0.29) signs 28999999 (0.28999999 ETH). Of the two-decimal amounts from 0.01 to 1000.00, 1,196 are affected at the USDC scale (1e6); of those from 0.01 to 100.00, 698 are affected at the 1e8 scale.

Why

lighter/signer_client.py (before this change): line 941 quote_amount = int(quote_amount * 1e6), lines 1238, 1353 and 1369 amount = int(amount * self.ASSET_TO_TICKER_SCALE[asset_id]), and line 1471 usdc_amount = int(usdc_amount * self.USDC_TICKER_SCALE). 2.01 * 1e6 == 2009999.9999999998 and 0.29 * 1e8 == 28999999.999999996; int() truncates both down.

How

A small to_ticks(amount, scale) helper scales through the amount's decimal representation (Decimal(str(amount)) * Decimal(scale)) and truncates the result, so 2.01 becomes exactly 2010000, while precision finer than one tick is still truncated as before (1.0000009 USDC stays 1000000). The five call sites use it. No other behaviour changes.

Testing

Added test/test_amount_scaling.py, which calls SignerClient.withdraw with a stub sign_withdraw and checks the amount it receives.

• Before the change: pytest test/test_amount_scaling.py → 2 failed (USDC 2.01 → 2009999, ETH 0.29 → 28999999), 1 passed.
• After the change: pytest test/ → 4 passed.
• flake8 . --select=E9,F63,F7,F82 (the CI's blocking lint) on the changed files → 0.