Off-by-one circuit index validation allows heap OOB in prover and verifier

[jilgolfo]

In lib/proto/circuit.h, CircuitRep::from_bytes() has off-by-one bounds checks for quad gate and wire indices:

// `lib/proto/circuit.h:215` – comment says "index of quad must be < wires in the layer"
if (g > max_g) return nullptr;   // accepts g == max_g (one-past-end)

// `lib/proto/circuit.h:220`
if (hl > nw || hr > nw) return nullptr;  // accepts hl/hr == nw (one-past-end)

lib/proto/circuit.h:224 in the same function uses the correct >= pattern for an analogous constant-index check (if (vi >= numconst)), confirming the inconsistency.

These boundary values are later dereferenced without bounds checks.

A malformed circuit blob with a single quad index at the boundary value (g == max_g or h == nw) is sufficient to trigger OOB on prover evaluation, and can also trigger OOB on verifier-side EQ indexing.

Fix: Change > to >= in both checks.

[matteo-frigo]

Thanks. I have fixed the bug in our Google internal code base, and the fix will be on github next time we push.