Restore the remaining missing diacritics in the Portuguese locale

[nank1ro]

Follow-up to #326, which only caught part of this.

That PR's detector looked for description prose of 120+ characters containing no diacritic at all. It therefore missed every section that had some accents alongside a few misspelled words, and it never looked at --instructions--, --answers-- or --solutions--.

A word-level detector finds the rest — 811 sections in 540 files:

word	occurrences
codigo	456
variavel	135
saida	131
nao	124
condicao	97
instrucao	66

Repo-wide the accented spellings dominate — "código" 1459 to "codigo" 540, "função" 2037 to "funcao" 26 — so the unaccented text is a straggler, not a house style.

What changed

The missing accents, and nothing else. No rewording, no reordering, no punctuation.

The judgement calls are the ones that change meaning rather than spelling: e (and) vs é (is), a vs à (crase), esta vs está, tem vs têm. Each was resolved by reading the clause — several lines contain both readings in sequence, e.g. "time é menor que 12, e no bloco de código imprima".

Diff shape

• 537 exercise files, 1064 insertions / 1064 deletions — line-for-line.
• 19 _theory.md files, +49 / −49.
• curriculum.json unchanged.

Checks

• Round-trip proof, 811 of 811. Stripping every combining mark from the new text (and mapping ç→c) reproduces the old text character-for-character. That is what proves only accents moved and no prose was quietly rewritten.
• Every changed exerciseType: 3 file still has its --solutions-- bullet character-identical to one of its --answers-- bullets — the validator matches those as exact strings.
• --output--, --asserts--, --seed-- and the other executed sections: byte-identical. This mattered: it's equivalent defect hides inside print("E' ora del caffe'!"), which is byte-compared against --output--, so accenting blindly would have broken exercises.
• Fenced code byte-identical everywhere except pt/python/dictionaries/2.md, where the change is a Portuguese prose comment (# obtem → # obtém). The code itself is untouched.
• Detector re-run: three apparent hits remain and all are correct as written — "Codigo" is a string literal, indices/lastIndex are Kotlin API names, and especifica there is the verb (no accent), not the adjective.
• validator: +275436: All tests passed!