[potatman]

Performance-only changes to the short-rate tree path (callable bond OAS / tree pricing). Every numeric result is bit-identical to develop. This is built on top of #330.

Changes

• TermStructureFittingParameter.NumericalImpl.value(): binary search instead of FindIndex with a capturing lambda. This lookup runs for every node in computeStatePrices and every stepback, so the linear scan made the tree build roughly cubic and allocated a closure per call. Exactness: IsEqual accepts at most double.Epsilon difference, so while times are strictly increasing (tracked in setvalue, reset in reset()) only the lower bound and its two neighbours can match; they are checked in index order, returning the same first match as FindIndex. Otherwise it falls back to a plain linear scan.
• Hot-path preconditions (TimeGrid.index, TreeLattice.partialRollback, TermStructure.checkRange, TrinomialTree construction): if (!cond) Utils.QL_FAIL(...) instead of QL_REQUIRE with a capturing lambda, which allocates its closure on every call even when the check passes. Exception types and messages are unchanged.
• TrinomialTree: branching lists presized to their node count.
• TreeLattice.stepback: removes the dead locals that repeated each probability/descendant lookup; summation order unchanged.
• DiscretizedCallableFixedRateBond: resolves each coupon/callability grid time once per lattice instead of calling TimeGrid.index at every step (lazily, in the original order, so failing lookups throw at the same point).

Measurements

Black-Karasinski, 100 steps, one quote = OAS + cleanPriceOAS at base, +1bp and -1bp; median of 3 runs, net10.0.

Bond	a	Time before	Time after	Alloc before	Alloc after
30y, monthly calls	0	350 ms	125 ms	188 MB	19 MB
20y, semiannual calls	0	48 ms	24 ms	34 MB	4 MB
10y, annual calls	0	31 ms	17 ms	23 MB	2.9 MB
30y, monthly calls	0.03	363 ms	133 ms	196 MB	19 MB

Verification

• Bit-identical outputs (BitConverter.DoubleToInt64Bits, byte-for-byte compare of the two runs): 1,584 values (OAS, semiannual OAS, clean price at the OAS and at +/-1bp, OASD, clean price at a fixed spread, NPV) over 11 bonds (callable, putable, non-callable, sinking, premium, deep discount), 2 curves, Black-Karasinski a=0 and a=0.03 and Hull-White, 25/50/100 steps.
• 492,196 fitting-parameter lookups on adversarial inputs (subnormals, +/-0, NaN, infinities, unsorted and duplicate times, reset after unsorted) return the same index as FindIndex.
• Failing preconditions throw the same exception types and messages.
• QLNet.Tests: 540 passed, 12 skipped, before and after.

A further ~1.7x is available by fitting the Black-Karasinski tree with an analytic derivative (NewtonSafe seeded from the previous step's theta instead of Brent over [-50, 50]), but that changes results around the 13th significant digit, so it is not part of this PR. Happy to send it separately if you want it.

🤖 Generated with Claude Code