[behdad]

Fractional CFF2 locations need separate normalization paths for outline/store instancing and runtime metrics. Match HarfBuzz's float normalization and half-unit rounding toward positive infinity; keep the metric path's intermediate 16.16 normalization before its 2.14 coordinates.

Use the shared unquantized float-region-scalar helper, double CFF operand products/accumulation, consumer-specific metric arithmetic, and a continuous bounds action on the common Type 2 interpreter. Full output retains per-blend rounding; bounds evaluate without operand rounding. Initialize outline coordinates directly instead of computing and discarding a second normalization result and clone.

Validation: 466 Skera tests with all features, the shared-reader scalar regression, Clippy with warnings denied, and Rust 1.85 all-target/all-feature checks. The current review's 90 full and 38 partial CFF comparisons pass, including decoded outlines, metrics, axis records, Private DICTs, and second-stage samples. Partial re-encoding retains its documented coefficient/F2Dot14 quantization limits.

Stacked on #2264 and preserves its reviewed nested-expression and performance fixes. Shared reader APIs provide the scalar evaluation; write-fonts supplies owned table models and assembly. No secondary generated table tree.

[behdad]

Review findings before changes:

• The separate outline/store and runtime-metric normalization paths are justified by HarfBuzz's arithmetic. hb_ot_var_normalize_* uses a 16.16 intermediate before conversion to 2.14; subset-plan normalization uses the outline path. HarfBuzz overrides roundf to round ties toward positive infinity, which is relevant when comparing the implementation with libc names.
• VariationScalars delegates region ratios to the shared read-fonts float helper and retains float accumulation for metrics. Continuous control bounds use the same Type 2 interpreter, an argument buffer reused across commands, and double coordinates. Full blend emission retains its separate rounding policy. The independent path-operator tests are useful coverage, rather than relying only on reference font bytes.
• One redundant computation remains: AxisPlan::new calls Skrifa's axes.location(...), clones those coordinates, overwrites every outline coordinate, and later recomputes the runtime metric coordinates explicitly. Remove that discarded normalization and allocate the outline-coordinate vector directly.
• The PR description predates the shared reader scalar helper and the earlier backports, so it needs to describe the remaining diff and current validation accurately. Partial re-encoding remains quantized; the sampled corpus is not proof of universal equivalence.

I will make the normalization cleanup, run the Skera suite and comparison matrices at this boundary, and update the description.

[behdad]

Pushed the reviewed parent and the precision cleanup through the stack. Axis planning now initializes the outline-coordinate buffer directly instead of computing and cloning a normalization result that was immediately overwritten. The reviewed boundary passes 466 Skera tests, the shared reader regression, Clippy and Rust 1.85 checks; the 90 full / 38 partial CFF comparisons retain their results.

[behdad]

Superseded by the rebuilt, consolidated instancing stack in #2416–#2423. MATH and name-table overrides remain separately covered by #2415 and #2411.