Rank Word-Translation Hints from Sentence Translations

An English sentence has one list of Spanish translation hints for each whitespace-delimited word occurrence. You are also given one bestTranslation and ordered alternativeTranslations. Every translation has the same number of whitespace-delimited tokens as the English sentence.
For each word position, rank its hints in two stages:
Lexicographically sort the hints.
Assign sentence priority. A hint matching the normalized token at that position in bestTranslation has priority 0. Otherwise, its priority is j + 1 for the first alternative translation at index j whose token matches. A hint appearing in no sentence has priority alternativeTranslations.length + 1. Sort by priority, using the stage-one lexicographic order to break ties.
Normalization lowercases a token and removes all leading and trailing ASCII punctuation characters.,!?;:. Punctuation inside a token is retained. Return the sorted hint list for every word position, preserving the original hint strings in the output.

Function
rankTranslationHints(sentence: String, hints: String[][], bestTranslation: String, alternativeTranslations: String[]) → String[][]

Examples
Example 1
sentence = "I eat apples."
hints = [["me","yo"],["comer","como"],["fruta","manzana","manzanas"]]
bestTranslation = "Yo como manzanas."
alternativeTranslations = ["Me comer fruta.","Yo comer manzana!"]
return = [["yo","me"],["como","comer"],["manzanas","fruta","manzana"]]
Best-sentence tokens come first. For the final word, fruta appears in the first alternative and manzana in the second. Trailing punctuation is removed for matching.
Example 2
sentence = "Hello world"
hints = [["hola","buenas"],["tierra","mundo"]]
bestTranslation = "¡Hola mundo!"
alternativeTranslations = []
return = [["buenas","hola"],["mundo","tierra"]]
The contract strips only the listed ASCII punctuation. The leading inverted exclamation mark in ¡Hola is retained, so neither first-position hint matches and lexical order decides that list. The trailing ASCII ! is stripped from mundo!.

Constraints
The English sentence contains between 1 and 100 whitespace-delimited tokens.
hints.length equals the English token count.
Each hint list contains between 1 and 50 non-empty strings.
bestTranslation and every alternative contain the same number of whitespace-delimited tokens as the English sentence.
0 <= alternativeTranslations.length <= 20.
Lexicographic comparison uses the original string values.

The trick is that the two-stage sort collapses into one comparator. Sort by priority first, then by original string. Lexicographic order as a tiebreak is the same as stage one, so you don't need two passes. For each position i, normalize the token from bestTranslation and each alternative at i. Build a map from normalized token to the lowest priority: 0 for best, j+1 for alternative j, using first match only. Hints not in the map get alternatives.length + 1. Pitfalls: normalizing the hints themselves (don't, the spec compares hints as given against normalized tokens, though you should check whether hints are lowercase in your cases), stripping inner punctuation (don't), and stripping non-ASCII marks like the inverted exclamation mark in Example 2 (don't). Output the original hint strings. Complexity is tiny at 100 positions and 50 hints. StealthCoder is the hedge if the comparator logic slips under time pressure.

If this hits your live OA and you blank, StealthCoder solves it in seconds, invisible to the proctor.