HandleSketchExtension deserializes the bare extension bytes instead of the reconstructed full sketch

This description is AI generated, mistakes can be made.

Intent: Use the received extension buffer rather than the snapshot+extension concatenation. Minisketch::Deserialize asserts GetSerializedSize() == obj.size(), so any extension message aborts the node. This sits between critical and high: it is a remote crash from p2p input, but only an outbound reconciliation peer can send it. It is labelled high.

Behaviour change: Original: rebuilds the 2x sketch from the stored initial sketch plus the extension and decodes. Mutant: the buffer is half the expected size, the assert in minisketch.h fires, and bitcoind aborts.

Trigger: We initiate reconciliation with an OUTBOUND_FULL_RECONCILIATION peer. Its initial SKETCH does not decode against our set (difference larger than capacity), so we send REQSKETCHEXT. The peer replies with a SKETCH carrying the high elements (any non-empty skdata). This happens to honest peers during normal operation, and a malicious outbound peer can force it at will.

Spec violation: BIP-330: the extension carries only the higher elements, and the initiator must combine them with the initial sketch to form the extended sketch.

Severity if undetected: high

Why tests miss it: src/test/txreconciliation_tests.cpp (HandleSketchTest stops at EXT_REQUESTED; no HandleSketch call in EXT_REQUESTED phase), test/functional/p2p_txrecon_initiator.py (run_test: 'TODO: Add more cases, potentially including also extensions'), no fuzz target references txreconciliation

Plausibility: skdata is the function parameter and working_skdata a local copy a few lines above, so picking the wrong one of two near-identical names is a classic slip.

Equivalence risk: None on the extension path: sizes always differ by the initial sketch length, which is non-zero because an empty initial sketch terminates early.