Prepare solmath 0.2.0 release

[DJBarker87]

Summary

• prepare the 0.2.0 release with a minimal default surface and explicit model feature flags
• add fully on-chain Kim Boundary Integration American pricing, exponential NIG pricing, continuous arithmetic-Asian/TWAP pricing, rainbow options, and checked public APIs
• remove the American closed-form and ROM paths so KBI has no live builder, uploaded operator, or off-chain pricing dependency
• harden fixed-point arithmetic, overflow behavior, approximation kernels, and feature isolation for Solana/SBF consumers
• expand validation, benchmark reproducibility, integration guidance, security documentation, and formal ULP proofs

American KBI results

Accuracy is absolute dollars at a $100 strike against QuantLib 1.41 QdFp accurateScheme.

Corpus	Leg	Median	P95	P99	Maximum
Production, 100K requested	Call	$0.0000156	$0.0006946	$0.0013619	$0.0027441
Production, 100K requested	Put	$0.0000160	$0.0007078	$0.0014208	$0.0032638
Adversarial, 10K requested	Call	$0.0000234	$0.0011672	$0.0017998	$0.0029456
Adversarial, 10K requested	Put	$0.0000279	$0.0013621	$0.0019915	$0.0027250

Deployed Agave measurements over 2,000 accepted quotes per leg:

Leg	Average math CU	P99 math CU	Maximum full instruction CU
Call	381,096	388,749	390,628
Put	371,876	388,879	390,786

The isolated KBI SBF artifact is 293,360 bytes, a 108,512-byte linked delta over the identical 184,848-byte Anchor baseline.

NIG results

• production maximum error: $0.0005653 per $100 notional
• adversarial maximum error: $0.0013966 per $100 notional
• deployed maximum: 382,441 full-instruction CU
• isolated SBF artifact: 311,464 bytes, a 126,616-byte linked delta

Arithmetic assurance

• 13/13 Kani harnesses pass with 545/545 bit-precise checks
• U256 carry assembly, long-division limb bounds, and Knuth-D3 correction bounds are verified compositionally
• both fp_sqrt paths now return the exact integer floor through proved Newton/bisection transitions and an exact fallback
• approximation certificates, critical-invariant tests, and independent reference corpora are included

Release validation

• cargo fmt --check
• strict Clippy correctness, suspicious, performance, and panic/unwrap gates
• no-default, default, and all-feature test matrices
• Rust 1.79 MSRV and thumbv7em-none-eabi no-std checks
• rustdoc warnings denied and examples checked
• exact packaged-crate tests: 344 tests and 23 doctests
• cargo publish --dry-run --allow-dirty
• final source archive: 59 files, 256,381 bytes, 9,859 bytes below the 260 KiB release cap

Review notes

• KBI and NIG are quote-only paths near the 400K CU ceiling; downstream programs must budget account, CPI, and token work separately.
• The Kani U256 result is compositional and does not claim monolithic verification of the complete Knuth multiply-subtract routine, compiler lowering, or SBF execution.
• Repository-only reference generators and large integration corpora are excluded from the published crate.