fix: match the format argument like an input unit

[MFA-G]

The problem

Input units are normalized before lookup — lower-cased, then retried without a trailing s:

else units = units.toLowerCase()
prevUnits = units = unit[units] || (units.endsWith('s') ? unit[units.slice(0, -1)] : undefined)

The format argument went through neither step:

return result && ((result / (unit[format] || 1)) * ...)

So a unit spelling that parses perfectly well as input silently fails as a format and falls through to the || 1 ms fallback:

parse('2MIN')          // => 120000   ✓ accepted as input
parse('2h', 'MIN')     // => 7200000  ✗ expected 120
parse('2h', 'mins')    // => 7200000  ✗ expected 120
parse('2h', 'hrs')     // => 7200000  ✗ expected 2
parse('2h', 'Hours')   // => 7200000  ✗ expected 2
parse('2h', 'seconds') // => 7200000  ✗ expected 7200

parse('2h', 'M') is the nastiest case. M is a legitimate minute spelling on the input side in any casing, so the caller gets 7200000 — a plausible number in the wrong unit, 60000× off, with nothing to indicate the format was ignored.

The plural forms matter in practice because 'seconds' and 'mins' are exactly what you'd carry over from the input strings this library is built to read (parse('10 seconds') works, the readme advertises 1hr 20mins).

The fix

Extract the existing input-side normalization into lookupUnit and use it on both sides, so one spelling rule governs the whole API:

const lookupUnit = (unit, name) => {
  name = String(name).toLowerCase()
  return unit[name] ?? (name.endsWith('s') ? unit[name.slice(0, -1)] : undefined)
}

This is behavior-preserving for everything that already resolved: exact lower-case names hit the first lookup unchanged, and an unrecognized format still falls back to ms via the existing || 1.

One deliberate detail: the input path used unit[units] || ..., which retries on 0 as well as undefined. lookupUnit uses ?? so a locale that legitimately defines a 0 unit isn't re-looked-up. No bundled locale has one, so nothing changes today — it just avoids a latent trap for custom locales.

Validation

• npm test — 102/102 pass (95 before, 7 added).
• Reverting index.js and keeping the new tests fails exactly the 6 assertions covering the bug, so they're not vacuous.
• The added case parse('2h', 'bogus') === 7200000 pins the ms fallback so this can't quietly turn into a throw later.

Also added one readme line next to the existing custom output format example, since the format argument's matching rules weren't documented.