ml: encode the editor's round-3 answers — rules 27 → 28, glossary v0.7.0 (two shipped guesses corrected)

.dev/STATE.md

@@ -7,6 +7,17 @@ Roadmap detail lives in the work-plan tracker **#257**, not here (PLAN.md predat
 
 ## In flight
 
+- **ml round-3 answers encoded, unreleased (2026-09-21, branch `ml-round3-answers-ml22`)** — the
+  editor answered lecture-python-programming.ml#22 on 2026-09-19. `ml` rules 27 → 28, glossary
+  v0.7.0. Two answers overturned a guess: the comma between `-ഉം` items is **always** there
+  (v0.29.0 shipped "phrases only"), and *draw* is not a light verb (left unencoded — both forms
+  are his). Confirmed: English plural + suffix for retained nouns (own rule + lint), *prefer*
+  English, *provide* → `നൽകുക` (both acceptable). Further Reading boundary settled as the
+  reference section (`D-2026-09-21-…`), which unblocks the code move under #260. **Open: the
+  rule-2 rewrite** around his everyday-speech test — held for a held-out judge run, not shipped
+  on his say-so alone. Round 4 (lecture-python-programming.ml#23) **will be regenerated at
+  `@v0` once this is released** (Matt, 2026-09-21; the editor has been asked to hold off) — see
+  the log entry `2026-09-21-ml-round3-answers`.
 - **v0.29.0 released 2026-09-18** (release PR #312 `a6fda54`; payload #311 `847ca9f`; §4a gate
   **completed**: 84/84 sync runs, 28/28 delivery + 28/28 `engineVersion: 0.29.0` verdicts per
   lane; `v0.29` = `v0` = `a6fda54`; GitHub release published; `@v0` smoke read-off on #312)

.dev/decisions/D-2026-09-18-ml-further-reading-lists-stay-english.md

@@ -1,3 +1,8 @@
+> **Boundary settled 2026-09-21** — the question this record leaves open (where the editor draws
+> the line) was answered on lecture-python-programming.ml#22: see
+> `D-2026-09-21-ml-further-reading-boundary-is-the-reference-section`. The decision below stands;
+> kept unedited.
+
 # Malayalam: Further Reading lists stay in English
 
 **Context**: In the third inline review round (lecture-python-programming.ml#13, `matplotlib`,

.dev/decisions/D-2026-09-21-ml-further-reading-boundary-is-the-reference-section.md

@@ -0,0 +1,26 @@
+# Malayalam: the Further Reading boundary is the outside-pointing reference section
+
+**Context**: `D-2026-09-18-ml-further-reading-lists-stay-english` accepted that the `ml` edition
+keeps its Further Reading bullets in English, and left one question open: where the editor of
+record draws the line — that section heading, any list of external links, or any bullet that
+opens with a link. It was put to him on lecture-python-programming.ml#22. He answered on
+2026-09-19: *"Keep the Further Reading section in English - this was intentional. Further
+Reading is reference material pointing outside the lecture - the reader will follow those links
+into English content anyway."*
+
+**Decision**: The boundary is the **section**: a closing block of reference material whose
+bullets point outside the lecture ("Further Reading", "References" and the like). The test is
+his reason, not the markup — the reader is about to land in English content, so an English
+pointer serves that reader better. It follows that the ruling does **not** extend to a link-led bullet
+in the body of a lecture, nor to a list that points at other lectures of the same edition
+(those land in Malayalam): both are ordinary prose and are translated. The 09-18 record stands
+unchanged; this one settles the question it left open, and the existing prompt rule already
+describes exactly this region, so its wording does not change.
+
+**Consequences**: The boundary now has no judgement in it for the heading-anchored case, so it
+can move from prompt to code as the 09-18 record anticipated — restore the list under a
+recognised reference heading from the source bytes, `verbatim-directives.ts`-style, and check it
+in `diff-checks`. That is follow-up work under #260, not part of this change; until then the
+prompt rule carries it (regeneration evidence: Further Reading verbatim 6/6 in the 2026-09-18
+arm). A reference list that sits under an unrecognised heading stays with the prompt rule.
+#promote

.dev/log/2026-09-21-ml-round3-answers.md

@@ -0,0 +1,34 @@
+# 2026-09-21 — ml round-3 answers (ml#22) encoded; two shipped guesses corrected
+
+**Trigger**: Adisankar answered QuantEcon/lecture-python-programming.ml#22 on 2026-09-19 — every question, one line each. Three answers confirm our reading, two overturn a guess, two are "both acceptable".
+
+| Item | His answer | Against our guess |
+|---|---|---|
+| Which ordinary words stay English | "would a Kerala student say this word in everyday conversation?" is "a fair test" | confirmed — mildly, and to a leading question |
+| *provide* | both acceptable; `നൽകുന്നു` read better in that lecture | soft confirmation |
+| plural of a retained noun | English plural + Malayalam suffix; agrees `-ുകൾ` after a hyphen renders badly | confirmed |
+| *prefer* | keep in English | we had none |
+| *draw* | `വരയ്ക്കാം` is natural and worked better there | **overturned** (`draw ചെയ്യുക`) |
+| comma between `-ഉം` items | always — phrases and single words | **overturned**, and v0.29.0 shipped the guess |
+| Further Reading | intentional: reference material pointing outside the lecture | the section, not any link-led bullet |
+| our two normalisations on ml#13 | both correct | — |
+
+**Done** (branch `ml-round3-answers-ml22`)
+- `ml` rules 27 → 28. The clause-boundary comma rule, part (c), loses "when each side is a phrase rather than a single word" and gains single-word examples; the suffix rule's own example (`a green border-ഉം a blinking cursor-ഉം`) contradicted it and now carries the comma. New short rule: the plural of a retained noun is the English plural plus the suffix — its own rule rather than one more clause in the longest rule of the set, where adherence is weakest.
+- Glossary v0.7.0: `prefer` pinned English (light-verb form); `provide` → `നൽകുക`, superseding the v0.4.0 pin, with a context note that the other form is never an error.
+- **Not encoded, on purpose**: *draw*. He changed `വരയ്ക്കുന്ന` to `draw ചെയ്യുക` at one line of `matplotlib` and calls `വരയ്ക്കാം` natural at another; both are his. The round-3 disposition report classed the first as a rule miss — in hindsight it was not. A future flip on *draw* or *provide* is preference, not a defect, and should be read that way in the round-4 analysis.
+- `D-2026-09-21-ml-further-reading-boundary-is-the-reference-section` (pointer note atop the 09-18 record). The prompt rule's wording already matched; moving it to code is now unblocked under #260.
+- `ml_metrics.py`: two lints — a Malayalam plural on a Latin-script singular, and two adjacent hyphen-suffixed `-ഉം` items with no comma. Calibrated on the three reviewed lectures and the round-4 draft: no false positives; they do find seven instances on already-merged pages (`matplotlib` 87, 89 ×2; `functions` 279; `python_by_example` 95, 358, 441), handled on the edition.
+- Tests: the two corrected strings pinned (and their old forms pinned absent), the plural rule, *draw* absent from the light-verb list, glossary shape. Suite green, lint/format clean, `dist/` + `dist-action/` rebuilt.
+
+**Still open — rule 2.** He accepts the everyday-speech test, so the rewrite the round-3 report proposed (the principle in the prompt, the word lists in the glossary) is unblocked. It is not in this change: rule 2 is the rule behind the ordinary-word pendulum (round 2: 24 to English / 8 to Malayalam; round 3: 4 / 12), *draw* — which we had listed under "stays English" — shows the test is not sharp, and nothing reaches him unmeasured. Run it held-out through the blind pairwise judge (`judge.mjs` in the 2026-09-18 arm) first.
+
+**Round 4 will be regenerated at the next release (Matt, 2026-09-21 — reverses the first call of the day).** The first call was to leave lecture-python-programming.ml#23 (`numpy`, v0.29.0) alone: about four contradicted lines in 143 did not justify three new draws. The arm below changed the arithmetic — the draws now exist, and branch draw 1 is cleaner than the draft he was sent on every lint (bare endings 0 vs 3, comma-less `-ഉം` pairs 0 vs 2, `provide ചെയ്യ…` 0 vs 6, banned 2 vs 3) — and his review time is the scarce resource. He has not started (no activity on #23 since 2026-09-18), and he has been asked on #22 and #23 to hold off. Sequence, as for round 3 at v0.28.0: merge → v0.29.1 with its §4a gate → three draws at `@v0`, best by lint, `ml_repair.py` applied and disclosed → force-update #23 and retitle. No branch draw goes to him: every round has been generated at a published `@v0`. If he says he has already started, #23 stays as it is and the pre-registration stands: flags on the comma-less pairs and any *provide* flip are known-and-encoded, excluded from the convergence residue.
+
+**Measured** (`experiments/ml-benchmark/arms/2026-09-21-round3-answers-numpy/`): three draws of `numpy` on the branch beside the three archived v0.29.0 draws — same source, model and command, and `numpy` is held out. The glossary change is deterministic (`provide ചെയ്യ…` 16 → 0). The corrected comma rule moves the output part-way only: hyphenated `-ഉം` pairs with a comma 0 of 4 → 2 of 6, the same two lines missed in both draws that had them. A rule he states as "always" lands about a third of the time from the prompt — the terminal-punctuation finding again — so `ml_repair.py` gained the comma as its third mechanical repair. The plural rule had nothing to do on this lecture; the one Malayalam plural in six draws is on a mass noun (`functionality-കൾ`), the same case as `logic-ുകളെ` on the reviewed `functions` page, and that is a question to him, not a rule.
+
+**Harness**: the local `dist-action` bundle delivered on the `.ml` lane (test-translation-sync.ml#191, fixture PR test-translation-sync#890): 87-term glossary loaded, UPDATE path, the plural rule firing on the fixture's own `concepts-കൾ` → `concepts`, the hortative holding. It also missed the comma on its one `-ഉം` pair and opened a sentence with lowercase `modern` — the second harness run in a row to do so. Both are lintable misses of rules already present: more weight on the lint-and-repair step (#260, #301) over more prompt.
+
+**Edition side**: lecture-python-programming.ml#24 brings the six places on the reviewed pages into line with his answers (three plurals, three commas), for his review; `logic-ുകളെ` left as his call.
+
+**Next**: review + merge → patch release v0.29.1 (nothing here is a feature) → regenerate round 4 at `@v0` → then the rule-2 arm.

CHANGELOG.md

@@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- **`ml` rules 27 → 28, glossary v0.6.0 → v0.7.0** from the editor's answers to the round-3 questions (lecture-python-programming.ml#22, 2026-09-19). Two answers overturned a guess that v0.29.0 had shipped or was about to: the comma between `-ഉം` items is there **always**, single words included (the clause-boundary rule said phrases only, and the suffix rule's own example contradicted it — both corrected), and *draw* is **not** forced into the light-verb pattern (he calls `വരയ്ക്കാം` natural; it stays unencoded either way). Confirmed and encoded: the plural of a retained English noun is the English plural plus the Malayalam suffix (`objects`, `function calls-ൽ` — never `object-ുകൾ`), as its own short rule; *prefer* stays English; *provide* moves to `നൽകുക`, superseding the v0.4.0 pin, with both forms acceptable in reviewed text. He accepts "would a Kerala student say this word in everyday conversation?" as the test for which ordinary words stay English; the rewrite of that rule is held for a held-out judge run. Decision record `D-2026-09-21-ml-further-reading-boundary-is-the-reference-section` settles the Further Reading boundary (the outside-pointing reference section, not any link-led bullet), which unblocks moving that rule into code.
+- **`ml_metrics.py` gains two lints, `ml_repair.py` a third repair**: a Malayalam plural built on a Latin-script singular, and two adjacent hyphen-suffixed `-ഉം` items with no comma — calibrated against the three reviewed lectures and the round-4 draft with no false positives. The comma is also repaired mechanically, because the measurement says it must be: on three held-out draws of `numpy` (`experiments/ml-benchmark/arms/2026-09-21-round3-answers-numpy/`) the corrected prompt rule lifts comma-carrying pairs only from 0 of 4 to 2 of 6, while the glossary change is deterministic (`provide ചെയ്യ…` 16 → 0).
+
 ## [0.29.0] - 2026-09-18
 
 ### Added