AlgoScan

The BlockHeader gained a Load field (codec "ld") and CongestionTax
("ct") in go-algorand-sdk. Bump the dependency so the algod importer
can decode current nightly blocks; otherwise GetBlock fails with
"no matching struct field found when decoding stream map with key ld".

The BlockHeader gained a Load field (codec "ld") and CongestionTax
("ct") in go-algorand-sdk. Bump the dependency so the algod importer
can decode current nightly blocks; otherwise GetBlock fails with
"no matching struct field found when decoding stream map with key ld".

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

We used to round (up) the fee obtained by multiplying usage by minfee
for each _group_.  This meant that the fee could be rounded up once
for ever group (top and inner).  Therefore, a simulate caller, having
obtained the usage for the group as a whole, could not really know how
much fee to add.

Now we carefully track our rounding so that after rounding up, the
excess (called "residue") is retained, and the next group can use that
residue to avoid rouding up.  The final result is that callers may
presume that there is exactly one rounding up for an entire group,
including inners.

I also removed the TotalUsage and TotalFeesPaid fields from the
simulation result.  These values were the sum across groups. (You can
simulate more than one top-level group at a time.)  But summing usage
across groups has the problem I described above. Instead, callers
should treat each group separately.

Per generated client crate, add tests/ (preserved across regeneration by the
src-only orphan sweep):
- models.rs: offline serde tests for the new extension-driven types
  (algokit_transact::Address, fixed [u8; N] arrays, box/holding/locals refs)
- integration_localnet.rs: #[ignore]'d smoke tests against a real localnet via
  the crates' .localnet() constructors, run with 'cargo test -- --ignored'

The unit tests run in the default offline suite; the integration scaffold is a
starting point for a fuller suite in a follow-up PR.

Instead of wiping the whole output directory before generation, overwrite the
generated files and sweep only orphaned .rs files under src/ (those no longer
produced by the current spec). Files outside the generated src/ tree -- e.g. a
hand-written tests/ directory -- are now preserved across regeneration.