Proposal: review each remaining lecture as a fresh draft on its own PR

[mmcky]

@Honaminto — thank you for your answers of 23 September on #77, #78, #80 and #42. They are all recorded below, under "What stays the same". This issue is a proposal about the 22 pages you have not reviewed yet, for you and @mmcky to settle together here. In short: instead of correcting the French that is on main now, you would review one lecture at a time as a fresh draft (a new translation made from the current English by the current version of the translation tool), opened as its own pull request. Carrying on as now is a perfectly good answer too; nothing changes unless you agree. (#40 says the translations are not regenerated during the review; this would change that, but only for the lectures you have not reviewed yet.)

What we propose

For each remaining lecture, when you are ready for it:

1. We make a fresh draft, check it (structure and code against the English, French typography, your rulings) and open it as its own PR. The PR replaces the lecture's current French, comes with a rendered preview, and says which version of the tool and which model made the draft, what we checked and anything we changed by script.
2. You review it on the PR in whatever way suits you: commit to the PR's branch directly (you now have write access), leave suggestion blocks, or write your numbered before/after list in a review comment, like the lists in your earlier PRs. We apply suggestions and lists for you, with you credited as co-author.
3. When your review is complete, approve the PR or say so on it, and leave the merge to us. (On Review of the French translation of Names and Namespaces #80 we said you could merge pull requests; for these drafts we would merge, because we update the tool at the same time.) We then go through your corrections, update the lecture and merge it.
4. Before the next lecture is drafted, we put what you corrected into the tool where it applies more widely (glossary entries, rules, automatic checks), so that a correction you make once is less likely to come back in the next lecture.
5. The merged lecture is kept as a reference for measuring later versions of the tool.

The Malayalam edition (QuantEcon/lecture-python-programming.ml) has been reviewed this way since August, and the new Japanese edition (QuantEcon/lecture-python-programming.ja, being set up now) will be built and reviewed this way from its first lecture.

Why

• The lectures waiting for you predate what the tool has learned from you. Apart from passages the sync has updated or added since (including the whole of polars, which the sync added at the end of July), and a few corrections by hand, they are the first draft made on 15 July. The French style rules we wrote from your first two merged PRs ([numpy] Corrections de style et de terminologie #24, [python_by_example] Corrections de style et de terminologie #25) and your glossary answers of 16 July all came later. Under the current way, what we learn from you improves future updates from the English, but reaches the text you are about to read only when we apply it by hand (as in Round-2 consistency pass: espace de nommage and renvoyer edition-wide, plus two hand-ports #81).
• Fresh drafts already get much of it right. We checked your round-2 corrections (Update functions.md #77, Update oop_intro.md #78, Review of the French translation of Names and Namespaces #80) against three fresh drafts of each of those lectures, made with the tool on 23 September, just before the glossary entries described below. In 14 of the 29 places where you corrected wording or typography, none of the three drafts repeated the mistake, and in 8 more only one or two of them did; in several places the drafts even used your own words (« Nous entendons par là », « Python ne peut pas déterminer si », « (de même) »). What the drafts still got wrong was mostly terminology. The two clearest cases became glossary entries on 23 September. Before them, the coin side "heads" came out wrong in 6 of 12 test drafts (as « pile » in 5), and the three drafts of names.md wrote « espace de noms » every time (206 times), never « espace de nommage ». With the new entries, neither mistake appeared in our test drafts.
• It works in Malayalam. There, the editor reviews one fresh draft per round, and the corrections go into the tool before the next. In rounds 2 to 4 the editor replied the same way (suggestion blocks), and the share of Malayalam prose lines changed in review fell from 97% in round 2 (Malayalam translation: functions — calibration review round 2 (v0.3 rules) lecture-python-programming.ml#7) to 67% in round 3 (Malayalam translation: matplotlib — calibration review round 3 (v0.27.0 engine) lecture-python-programming.ml#13) and 51% in round 4 (Malayalam translation: numpy — calibration review round 4 (v0.29.2 engine) lecture-python-programming.ml#23). The lectures differ, so the figures are only a rough guide, but the round-4 review summary was "translation improved exponentially".
• Each lecture you finish becomes a test case. Because each draft comes from a known, current version of the tool, your corrections show what that version still gets wrong, not what a July version got wrong and has often learned since. Once merged, the lecture is a permanent reference: we can translate the same English with later versions and measure how close they come to your text, so we know whether a change to the tool really helps the French.

What would change for you

• You would review a PR that replaces the lecture's current French, rather than opening your own PR against it. The "Files changed" view will show most of the prose as changed (about two lines in three in our test drafts), but lines that the new draft words exactly as before show as unchanged, and some are folded out of view, so please read the whole lecture in the preview. Suggestions work on any line the view shows; for any other line, a plain comment on the PR is fine.
• Every PR has a rendered preview. None of your seven PRs so far had one, because they came from your fork (with write access, a branch of your own in this repository would now get one too).
• Automatic reviews (GitHub's Copilot, and our own translation check) may comment on these PRs. They need no attention from you.
• We would open each lecture when you are ready for it, not all 22 at once. One at a time is slower than having several lectures open: between lectures there is a gap of some days while we put your corrections into the tool and make the next draft.
• Each lecture's item on French translation review — native-speaker read-through (project tracker) #40 still completes when that lecture's PR merges.

What stays the same

• The five lectures you have reviewed (numpy, python_by_example, functions, oop_intro and names) keep your text. They are not regenerated; we only bring them in line with your answers of 23 September (for example « fonction native » and « cours »).
• The sync from the English keeps running for the whole edition.
• Your rulings stand, including your answers of 23 September, which go into the tool before the first draft:
  • the imperative for exercise instructions, and « On pose » / « On suppose » for set-ups and assumptions (Update functions.md #77);
  • « fonction native », « type natif » and « espace de nommage contenant les primitives » (Review of the French translation of Names and Namespaces #80);
  • « paquet », « bibliothèque » and « gestionnaire de paquets », as you distinguished them (Terminology: seven glossary rulings (from action-translation#78 + cours/chapitre) #42);
  • « cours » for a lecture (Terminology: seven glossary rulings (from action-translation#78 + cours/chapitre) #42), which is already what the tool uses.
• There is no deadline, and the priorities and the list of what to ignore in French translation review — native-speaker read-through (project tracker) #40 are unchanged.

Costs and open points

• A fresh draft replaces everything in that lecture: the July text, passages the sync has updated since, and corrections we made by hand (for example in Round-2 consistency pass: espace de nommage and renvoyer edition-wide, plus two hand-ports #81). That is fine for a lecture nobody has reviewed, because the draft is translated from the current English and so already includes the English changes; we check each draft against your rulings before opening it.
• Terms the glossary does not cover can come out differently from one draft to the next. When one reads wrong, or differs from a lecture you have reviewed, say so on the PR and it goes into the glossary before the next draft.
• Index entries. You said on Update oop_intro.md #78 that the index entries should be in French throughout. The tool still writes them in English, so until it can, we would translate them in each draft before it reaches you, and we will still make the one pass over the whole edition that we agreed on, so the index is consistent in the meantime.
• More than one draft. We may make more than one draft of a lecture and open the cleanest; the PR would say so.
• If the English changes a lecture while its PR is open, we bring the PR up to date. There is nothing for you to do.

Questions for you

1. Would you prefer to review the remaining lectures this way, or to carry on as now?
2. Any concerns, for example about reading a new draft rather than the current text, or about the pace?
3. Which lecture next? Your fork has a review/python-oop branch, so perhaps python_oop.md (Review: python_oop.md #51). If you have already started on it, you could finish it the current way and start the new way with the lecture after; if not, it could be the first fresh draft.

If the proposal is adopted, @mmcky then brings #40 and the open lecture items in line with it, as a separate follow-up.

Done when

• @Honaminto and @mmcky agree here on how the remaining lectures are reviewed, and the outcome is recorded in a comment on this issue.