[Rajvira10]

With a descending domain whose first interval is collapsed, scaleDiverging().domain([1, 1, 0])(0) returns 0.5 instead of 1. Determine the direction from the transformed domain endpoints rather than the first interval, including when a power or symlog transform is reconfigured.

Fixes #256.

Regression coverage includes the reported inputs, clamping, copying, a collapsed last interval, power and symlog reconfiguration, and positive and negative logarithmic domains. All four new tests fail against the original source and pass with this change.

Validation on Node 26.8.1:

• Full Mocha suite: 437 passing, using the runtime dependency versions from yarn.lock (d3-array 3.0.4, d3-format 3.0.1, d3-time 3.0.0).
• ESLint across src and test: zero errors or warnings.
• git diff --check: passed.

The installed Mocha 9 CLI's yargs dependency is incompatible with Node 26, so the suite was run through Mocha's programmatic API, with the same test glob and timezone as npm test:

TZ=America/Los_Angeles node --input-type=module -e 'import Mocha from "mocha"; import glob from "glob"; const mocha = new Mocha({reporter:"dot"}); for (const file of glob.sync("test/**/*-test.js")) mocha.addFile(file); await mocha.loadFilesAsync(); mocha.run(f => {process.exitCode = f ? 1 : 0;});'
node --input-type=module -e 'import eslint from "eslint"; const cli = new eslint.ESLint(); const results = await cli.lintFiles(["src", "test"]); console.log(await (await cli.loadFormatter("stylish")).format(results)); const errors=results.reduce((n,r) => n+r.errorCount+r.warningCount,0); process.exitCode=errors?1:0;'

The diff was independently reviewed and tested.