By the third century BCE…
…probably. We're not quite sure when.
In a question about Old Latin meters, an anonymous user brought up Mercado's convincing argument that the Saturnian was based on accent. The idea isn't new, but Mercado backs it up with some nice information-theoretical analysis: basically, the accent theory passes some statistical tests that all the quantity theories I've seen fail at.
His idea is that a Saturnian line generally resembles one of the following:
'⏑'⏑|⏑'⏑ || ⏑'⏑|⏑'⏑
'⏑'⏑|⏑'⏑ || '⏑⏑|⏑'⏑
' marks a stressed syllable and
⏑ an unstressed one. For example, the Epitaph of Naevius:
Ímmortálēs mortálēs / si fóret fas flére,
flérent dévae Caménae / Náevium poétam:
ítaque póstqu' est Órchi / tráditus thesáuro
oblítī sunt Rómae / lóquier línguā Latínā.
If gods could weep for mortals, the divine Muses would mourn the poet Naevius: after he was delivered to the vault of Hades, Rome forgot how to speak Latin.
Mercado concludes that this pattern fits the majority of the surviving Saturnians, with some additional rules about elision and hiatus and two light syllables next to each other. (Unfortunately, Mahoney doesn't quote that part in her review, and the only copy of this book at the university library has been overdue for a year now—but I'll add more detail if I find a copy.)
Crucially, though, the Saturnian relies on the accent system of Plautus's era:
- If the second-to-last syllable is long, it gets stressed
- Else, if the third-to-last syllable is long, it gets stressed
- Else, if the fourth-to-last syllable is short, it gets stressed
- Else, the third-to-last syllable gets stressed
- Finally, every second syllable before that one gets lightly stressed, if applicable
This is notably not the Classical penult system. But it's very close: the difference is, the Classical system never lets the fourth-to-last syllable get the stress. This is a pretty rare edge case, showing up mostly in inflected forms like géneribus/genéribus, so I don't have qualms about calling this an early form of the "penult rule". (The "proto-penult rule"?)
A date of somewhere before the third century also makes intuitive sense. As Mercado puts it:
It would otherwise have to be assumed that the (pre-?) Old Latin [initial stress] system was replaced wholesale by the Plautine [proto-penult stress] one at almost the same time that Andronicus, Naevius, Plautus, and Ennius flourished.
Mahoney adds that:
…even if Saturnians were being composed when Latin had word-initial stress, the poems that survive presumably use the stress patterns of their authors' own time.
We know from other Italic languages that the word-initial stress arose after Venetic split off, but while all the other Italic languages were still a more-or-less-homogeneous blob in the middle of Italy. Etruscan had the same system, so Italic might have borrowed it from there. The timeline of these events is extremely fuzzy, but it seems like there was plenty of time for word-initial stress to leave its mark and then give way to the proto-penult rule.
So, short answer:
Carthágō delénda est.
(Or, for the purists: céterum cénseō Cartháginem ésse deléndam. Either way, Cato the Elder would have known the penult rule.)