+ xBRZ 1.9 as its own scaler entry

[Hawkynt]

Implements Hawkynt/2dimagefilter#30 from the derived specification merged in #22. Rung 2 throughout: the specification was the input, none of Zenju's code was read into this.

What 1.9 actually changed

Two things, both exact and integer:

1. Steep direction threshold 2.2 → 2.4, encoded here as the integer ratio x * 22 <= y * 10 becoming * 24. Ten sites — the shallow and the steep test in each of the five scale kernels. Fewer near-diagonal runs now count as lines.
2. Weighted and 50/50 blends round to nearest instead of truncating, so blended pixels stop being biased downwards by up to one step per channel.

What turned out not to be a gap

Worth stating, because all three look like gaps:

Suspected	Verdict
centerDirectionBias (1.7)	no — 1.7 turned a hard-coded 4 into a config field defaulting to 4. This port already uses 4.
Asymmetric colour distance (1.7)	no — an artefact of packing channel differences into a byte-indexed lookup table. This port calls a colour metric directly and never had it.
ARGB outside area transparent (1.8)	no — OutOfBoundsMode already exposes this per axis, configurably.

So the existing xBRZ entry was 1.7-equivalent, and 1.5/1.6 changed only API and performance.

The design question #22 left open

ILerp is picked by the pipeline — Color4BLerpInt<Bgra8888> is hard-coded at every shared call site — so the 1.9 path cannot select a rounding lerp type. Xbrz.InvokeKernel implements IRescaler, so its constraints are fixed by that interface, and it cannot substitute one itself either: TWork is generic there, so it cannot know whether the 3- or the 4-component rounding struct is right. Rounding inside the kernels was the other option and means not calling lerp at all — open-coding the channel arithmetic and discarding the colour-space abstraction.

So ILerp<T> grew a rounding counterpart method instead, LerpRounded, in both overloads. The kernels keep blending through the interface and call the member their variant calls for.

The members are declared, not defaulted. Default interface methods would have kept the interface source-compatible for outside implementors, but this package targets down to net35, and net35–net48 have no runtime support — the compiler rejects them outright with CS8701. BatchDistanceDefaults already exists here for exactly this reason. All eleven shipped implementations implement them explicitly, which also keeps them off the boxing path; these are hot loops reached through a struct generic constraint. Float-based and no-op blenders forward to the truncating member, with a comment saying why.

Outside implementors of ILerp<T> must add the two members. There is no alternative on these frameworks; the members are documented with the behaviour a forwarding implementation reproduces.

Registration

Xbrz19 is its own [ScalerInfo("xBRZ 1.9")] entry, following XBR / XBR NoBlend / XBR 3x Original. The existing xBRZ entry keeps its name, its public constructor and its pixels — a version selector that quietly changed the default would be a regression for every script and saved setting naming it. The variant rides as a struct type parameter on the per-scale kernels, so both knobs fold to constants on specialisation and the pre-1.9 path takes no branch.

Verification

All three properties are exact, so all three are pinned exactly.

• The existing entry does not move — output at all five scales pinned by SHA-256 against hashes captured from the code before this branch existed. Stable across net45, net48, netcoreapp3.1, net9.0, x86 and x64.
• Threshold — a source whose accumulated distance ratios straddle 2.2 and 2.4, compared between the two thresholds with blending held at truncating, so a difference can only be the threshold's.
• Rounding — blends whose weighted sum leaves a non-zero remainder against the total weight, covering the weighted and the 50/50 path, plus a kernel-level comparison with the threshold held at 2.2.

Two internal single-change variants make that attribution possible; asserting only against combined 1.9 would let either change go missing unnoticed while the other still produced a difference.

Every new assertion was confirmed to bite by perturbing what it tests — restoring a 22 at one of the ten sites, dropping a rounding term, and pointing the public entry at 1.9 — and checking the matching test fails, and only the matching test.

1901 → 1936 tests, green. 2dimagefilter builds against this branch and goes 665 → 667, green.