Unary minus on a BigInt array element float-negates it (wrong value, silent)

[proggeramlug]

Found while benchmarking BigInt division (#9141); reproduces on pristine main 653e886699, unrelated to that change.

Repro

const c = [2n ** 64n];
console.log(-c[0]);

	result
node	-18446744073709551616
perry	-18446744073709552000

The perry value is the f64 rounding of 2^64 — the BigInt is being converted to a double and float-negated. Silent wrong value, no throw.

Characterization

• Reading c[0] alone is correct (prints the exact BigInt).
• -v where v is a plain local holding the same BigInt is correct.
• Only unary minus applied directly to an array element mis-lowers.
• An explicit bigint[] annotation does not help.

So the element read produces the right value and the unary-minus lowering picks the numeric path regardless of the operand's runtime type.

Why it matters

BigInt arrays are the normal representation for big-number tables (hash constants, crypto limbs, byte→BigInt lookup tables like TypeBox's). A silently-rounded negation in that code produces wrong results with no diagnostic.

Repro file: secret-tests/scratchpad/probe_q2.ts.