index: IR staging and index read dominate non-engine build time, and AXIOMCODE_REINDEX=1 does not force a rebuild #1975

Part of #1835.

Two problems in axiomcode index on medium and large repositories:

1. Build time spent outside the engine. On a ~1,300-file Python project the IR staging step alone takes about 12 s of CPU, and the index step that reads the IR into graph.sqlite takes about 13 s wall / 10 s CPU. Neither does any resolution work; both are dominated by reading and rewriting text rows. A cold index of that project spends about 119 CPU-s in total.
2. AXIOMCODE_REINDEX=1 does not force a rebuild. It is documented to rebuild the graph, but when the source files are unchanged the build still answers "graph up to date" and rebuilds nothing.

Expected: staging and indexing cost a small fraction of the build, and AXIOMCODE_REINDEX=1 axiomcode index always rebuilds. The resulting graph must be identical.