fix(transforms): avoid NaN at CurveInterpolator's lowest sample point

[mairas]

Bug

CurveInterpolator::set(x) returns NaN when x equals the smallest sample's input. The interval search starts at begin() and only advances while input > it->input_, so at the lowest point it never advances: x0 and x1 both reference that same sample, and the interpolation (…)/(x1 - x0) divides by zero.

Fix

Guard the zero-width interval: when x1 == x0, output the sample's y directly.

How it was found

Surfaced by executing the test suite on a HALMET device (test_curve_interpolator_exact_sample_points — "Expected 0 Was nan"). CI only compiles the Unity suites (pio test --without-testing), so it never caught this. The existing test in test_transforms covers the fix; it passes on-device after this change.

[mairas]

Automated review

Verdict: Correct, minimal fix. Approve. Root cause and fix confirmed by source-tracing; no blocking issues.

Root cause — confirmed

samples_ is a std::set<Sample> ordered by input_ alone. In set(), after the input < x0 early return, the search loop (lines 66–74) advances it only while input > it->input_. At the lowest sample point input == begin()->input_, so the loop never advances: it stays at begin(), x1 = it->input_ equals x0 = begin()->input_, and the original (…)/(x1 - x0) was 0/0 → NaN. Confirmed.

Fix correctness — confirmed for all three paths

• Lowest sample point: x1 == x0, y1 == y0 (same sample); fix outputs y1 = that sample's y. Correct (returns the sample's own y, not a hardcoded 0).
• Interior exact hit (e.g. query x=2 on samples 1,2,3): loop consumes x=1 into x0, breaks at x=2, so x1 != x0 and the normal path yields y1. Unchanged by the fix.
• Between points: x1 != x0, normal path, unchanged.

x1 == x0 is reachable only at begin() — any later exact hit has already advanced x0 to the previous sample.

Degenerate paths — checked

• Duplicate sample x: Cannot occur. The set's ordering key is input_ only, so std::set treats equal-x samples as equivalent and rejects the second insert. No mid-table x1 == x0. The y1 choice is moot for duplicates and correct at begin().
• Float ==: Appropriate here. In the trigger case x0 and x1 are both copied from the same stored it->input_ (the iterator never advanced), so they are bit-identical; == is exact and no epsilon is warranted. The guard also doubles as a general divide-by-zero guard for any bit-identical pair.

Style / conventions

Minimal, idiomatic, and the comment explains the why (zero-width interval at begin()) rather than narrating the change — matches AGENTS.md comment guidance. No nits.

Non-blocking note

test_curve_interpolator_exact_sample_points asserts the lowest point returns 0.0, but that sample's y is itself 0 — so the test does not distinguish "returns the sample's y" from "returns 0". The fix is correct by source-tracing regardless; if you want the test to pin the behavior, a fixture whose lowest sample has a nonzero y (e.g. (0, 5)) would make the output_ = y1 choice explicit. Optional.