Although the compositional trends of bulk rocks and melt inclusions largely overlap, there is a prominent contrast in the petrographic distribution and relative frequency of melt inclusions between different rocks (Fig.  8 ) and between the compositions of melt inclusions in apparently coexisting phenocryst minerals in any one rock (Fig.  9 ).

These contrasts between melt compositions enclosed in mafic and felsic phenocrysts become even more obvious when average melt inclusions and bulk rock compositions are compared for individual rocks (Fig.  9 ).

Fractional crystallization with physical separation of crystals from the melt would generate curved trends (e.g., in Al ₂ O ₃ or K ₂ O) depending on the stable phases and cannot explain the widespread observation of highly contrasting melt inclusion compositions in andesites with low phenocryst contents (Fig.  9 ).

More generally, melt inclusions have highly contrasting compositions in different phenocrysts from the same rock (Fig.  9 ) and, thus, phenocrysts do not represent stable equilibrium assemblages.

For all the other rocks, the contrast in melt inclusion compositions (Fig.  9 ) suggests that phenocrysts formed in different melts and were mechanically mixed to generate the observed phase assemblage..

Textural evidence and chemically similar melt inclusions show that pyroxene is crystallizing together with plagioclase from these melts (Fig.  9 , sample NB5B; see also Fig. 6 in Halter et al.

In rocks where pyroxene and plagioclase co-exist, amphibole is partially resorbed (Fig.  9 ), recording a shift from a plagioclase (±amphibole) stable assemblage to the plagioclase+pyroxene stability field.

Melt inclusions recording this event show enrichment in K ₂ O, consistent with crystallization of potassium-free plagioclase and pyroxene at the expense of amphibole (Fig.  9 ).