#6 (perf/clash): SparseProof warmer variant ran full root-to-leaf proof reads
and threw away the result. Until the M5 background sparse task can consume
those reads and reveal them into the trie, the variant only warms DB/OS page
cache while burning CPU on RLP decode that nothing else benefits from. The
default warmer stays Legacy — no behavior change for the default config —
but the IFlatDbConfig doc comment and the SparseTrieWarmerVariant enum now
mark SparseProof as EXPERIMENTAL with the cost spelled out. Same for the
in-code comments at the two warmer call sites (FlatWorldStateScope and
FlatStorageTree). This way no operator picks SparseProof expecting it to
prefetch into the sparse trie.

#7 (arch gap): SparseTrieTask and SparseTrieProofPrewarmer were the M4/M5
background pipeline placeholders — they had real implementations but DI
never resolved them. Grepping the repo for SparseTrieTask, SparseTrieProofPrewarmer,
HashedStateUpdate, or ProofResult returns nothing outside the files themselves.
Deleted both files. Wiring the real pipeline is the M4/M5 milestone work and
will reintroduce equivalent types with the actual DI registration; keeping
them around now mostly mislead readers into thinking the background sparse
task was active when commit/root work is still synchronous.

#8 (test gap): BlindedBoundaryProofTests already covers UpdateLeaf via
blinded boundary, ExtensionOnlyReveal-then-update, and MultiBlockReuse
across the blinded path. Attempted to add a synthetic dup-RLP-different-path
test for the (path, rlp) grouping fix but the proof walker validates the
RLP-by-hash against the path, so a path-agnostic mock reader would be
needed and the construction adds more confusion than coverage. The
production-derived UsdtMainnetReproducer in State.Flat.Test already
exercises the multi-target / dup-content blinded scenarios end-to-end.