What weak SE adds to KS. In knowledge soundness, the adversary gets only the CRS. In weak simulation-extractability, the adversary additionally queries a simulation oracle q times, obtaining simulated proofs on statements φk of its choice (possibly false). It wins if it outputs a verifying proof on some statement φ* that was never queried, from which no witness can be extracted.
Why this is hard. Each simulated proof adds three new group elements the AGM adversary can algebraically combine into its forgery. The risk is that A splices pieces of several simulated proofs together — perhaps one μ from query 1, one ν from query 2, etc. — to fabricate a valid proof on a fresh statement.
The payoff. This walkthrough shows the splicing attack is impossible. Of the five new coefficient families A8,i, A9,i, B5,i, C8,i, C9,i the AGM adversary could introduce, eleven carefully-chosen monomial coefficients from the verifier equation force a sharp dichotomy: either all simulation coefficients vanish (→ KS extractor recovers the witness), or exactly one index k survives (→ the β-critical equation forces φ* = φk, contradicting the winning condition).
Assumption in force. Two mild QAP conditions: the instance polynomials {ui(x)}ℓi=0 are linearly independent, and their span is disjoint from the span of the witness polynomials {ui(x)}mi=ℓ+1. Both are satisfied by any sensibly-constructed QAP.