ml: split comma-joined clauses in seeds; lint the whole lecture (round 4)

[mmcky]

Summary

This PR updates the Malayalam calibration tooling from the editor's fourth review round (QuantEcon/lecture-python-programming.ml#23, numpy, 73 suggestion blocks, merged 2026-09-28). Two changes, both in experiments/ml-benchmark/scripts/:

1. ml_repair.py gains a fourth deterministic repair: it splits two finite clauses joined by a comma and a resumptive pronoun or connective into two sentences (… methods ഉണ്ട്, ഇവയെല്ലാം … → … methods ഉണ്ട്. ഇവയെല്ലാം …).
2. ml_metrics.py now sees the whole lecture, and its line-level lints stop firing on hard wraps.

Nothing in src/ or dist-action/ changes, and CI does not cover these scripts. This is the prototype stage of #260 (post-processing), feeding #301 (lint gaps).

Why

The largest class in the round was comma-joined clauses where the editor writes a full stop: 20 flags, 23 comma joins removed.

	Splice sites per 100 Malayalam prose lines
Nine numpy draws (v0.29.0–v0.29.2)	13.1–21.0 (mean 17.2)
His reviewed pages, rounds 2–4 (functions, matplotlib, numpy)	1.6–2.8
His round-1 page (python_by_example), before he was splitting these	8.7

Rule 12's "split … at comma splices" has been in the prompt throughout. Most sites don't come from an English comma splice: the model renders an English relative clause, so that, a participle or an appositive as comma + resumptive pronoun. A prompt rule does not reach this, and the programme's measurements say "always" rules land about a third of the time anyway, so this goes to lint + deterministic repair.

What changes

The repair (comma_splice_resumptive lint → ml_repair.py)

• Match: a finite verb ending, then , , then a word from an explicit list — ഇത്, ഇതിനെ, അത്, അതിനെ, ഇവ/അവ with enumerated case endings, അതിനാൽ, അതേസമയം — with the clause carrying on after it on the same line.
• Change: the comma becomes a full stop. No comma is ever added after the connective.
• Guards:
  • the finite token must not open its sentence;
  • എല്ലാം ("all", which ends like the modal -ാം) and the fronted ശ്രദ്ധിക്കുക/ഓർക്കുക that rules 13/14 require are skipped;
  • a pronoun that ends the line (it opens a list) is not split, since that would leave a verbless അത്:;
  • the neither-nor -ഉം അല്ല, and a *അല്ല*, ഇത് … contrast are skipped;
  • a trailing "as shown" tag and a list-final , ഒപ്പം are skipped;
  • അങ്ങനെ, അതുകൊണ്ട് and പക്ഷേ are deliberately excluded: he writes , അങ്ങനെ himself.
• One classifier: ml_repair calls ml_metrics.split_resumptive_splices, so the lint and the repair cannot disagree.

The report-only watch

• comma_splice_watch lists every comma-joined finite clause, and comma_splice_rate gives the per-100-line figure.
• His reviewed pages keep one to four per lecture, so zero is not the target. Use it as a secondary best-of-N key at most.

Prose scanner fixes (prose_line_info, with prose_lines kept as a wrapper)

• A paragraph opening with ( was skipped as if it were a (label)= line. Five numpy lines were invisible to every lint, including a bracketed paragraph with no stop that is bare in 9/9 draws. Only whole-line labels are skipped now.
• {note} and other prose-directive bodies are scanned (directive names case-insensitively). One of the round's splices sat in a {note}. Exercise-family directives, figures and code are never scanned. Exercise content is English by decision (D-2026-09-03-ml-all-exercise-content-stays-english, enforced by src/verbatim-directives.ts). As a consequence, python_by_example's round-1 exercise text, translated before that decision, is no longer linted; one hit is lost there (seed line 561, which he fixed).
• Fences follow CommonMark. The old toggle treated any line starting with three backticks as a closer. It lost parity at a nested {hint} inside an unclosed {exercise-start} in python_by_example, then linted a code cell as prose and skipped the prose after it. A fence body now ends only at a bare closer at least as long as its opener.
• Terminal punctuation is checked at paragraph ends, outside list items (wrapped continuations included), and capitalisation at paragraph starts. This removes the hard-wrap false positives: 11 fewer terminal hits and 3 fewer lowercase hits on the round-1 seed.
• A closing bracket counts as terminated only after a stop, in the lint and the colon repair alike. (….) and ...) pass; (… memory) is flagged. He added that stop at ml#23 line 197 and in round 2, and accepted ...). The deep-copy line before a cell now gets his accepted ): (ml#23 line 920); it was bare in 7 of 9 draws and the repair used to skip it. A wholly bracketed paragraph is left to the lint.
• A comma counts as an ending only before a list it opens, not before a code or math cell. That case is flagged and left to the lint, since the repair would write ,:. It occurs in no ml corpus. This follows Copilot's review.

Retired: the future-hortative watch. Since rule 13 landed it has made two hits on reviewed seeds and the editor acted on neither. It flagged a later-lecture promise he kept, and missed another.

Validation

Repair, round-4 seed:

• It splits exactly the 13 comma-plus-pronoun sites, all on lines he flagged.
• His text has the same full stop at 10; the other 3 he folded into a converb.
• It reproduces three of his lines byte for byte (ml#23 lines 336, 494 and 841, used as fixtures).
• It fires on 0 of the 70 lines he left untouched.

Repair, earlier rounds:

• In the round-1 and round-2 seeds he removed the comma at all 12 such sites he edited: a full stop at 9, an em-dash at 2, a semicolon at 1. A 13th sits in a solution he reverted to English. Across all three rounds the full stop is his form at 19 of the 25 edited sites.
• It makes no change to any of his four reviewed pages (python_by_example, functions, matplotlib, numpy).
• On the two unsent current-engine draws it fires 12 and 15 times, all in-class on inspection.

Lint changes on his reviewed pages: hits are only removed, apart from two new true positives on python_by_example lines the old parser never saw: a lowercase short opener, and thought-ഉം planning-ഉം, which is the pair the editor's "always" answer on ml#22 covers.

Coverage diff: 38 files (the ml edition, every ml arm draw, the review seeds). Every added prose line is a ( paragraph, a prose-directive body, or prose the old toggle had lost after a parity error. Every Malayalam line no longer scanned sits inside an exercise-family directive, and those stay English by decision record.

Tests: python3 -m unittest discover -s experiments/ml-benchmark/scripts -p 'test_*.py' runs 16 stdlib tests. They cover his exact lines, one fixture per guard (including his reviewed python_by_example 275, which only the എല്ലാം guard protects), the parser cases (the {exercise-start}/{hint} desync, a code cell nested in a longer-fenced {Note}) and the bracket rule. Each of twelve single-guard mutations (removing a guard or reverting a regex) fails the suite.

Adversarial review before opening. Four reviewers ran the change over every ml corpus: parser, splice precision, lint regressions, and claims. Each finding went to two sceptics who reproduced it. 15 findings; 8 were confirmed by both, all low severity once verified, and all 8 are addressed in this PR:

• a verbless split at a pronoun that ends a line (2 of 285 sites, both in the oldest draw);
• the colon repair disagreeing with the new bracket rule;
• wrapped list-item continuations flagged as bare endings;
• overstated rounds-1–2 and reference-band figures;
• tests that did not reach the guards.

None of the 15 findings showed the repair touching his reviewed text. Of the rejected findings, the exercise-body scope is kept on purpose (see above). The rest were either intended behaviour or inputs that do not occur in any corpus.

Follow-ups (not in this PR)

• A repair for bracketed paragraphs with no inner stop, and a generalised -ഉം comma (negated pairs, multi-word items, chains).
• The P2 lexical repairs from the round:
  • imperative see → നോക്കുക
  • adverbial element-wise → ഓരോ element-ലും
  • have already seen → കണ്ടുകഴിഞ്ഞു
  • standalone മുൻ → മുൻപത്തെ
• Promotion into the engine under W2 — localisation becomes deterministic (after D1) [P1] #260, at model output only (file-processor.ts:689), never in the document-wide typography pass at :559. That pass would touch editor-reviewed sections.

The full per-flag analysis is in QuantEcon/project-translation reports/2026-09-28-ml-numpy-review-disposition.md.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🟡 Changes recommended

The terminal-punctuation lint currently treats a trailing comma as acceptable even before code-cell/math blocks, which contradicts the documented convention and can suppress intended linting.

Review effort: Lite
Findings: 1 · 1

Open (2)

• Trailing comma wrongly suppresses code-block terminal punctuation · New
• Broken docstring link splits repo name and anchor · New

What changed in this PR

This PR updates the Malayalam benchmarking/calibration scripts under experiments/ml-benchmark/scripts/ to (1) add a deterministic “comma splice + resumptive” sentence-splitting repair and (2) improve prose scanning so linting operates over the full lecture while avoiding hard-wrap false positives. It also adds fixture-driven stdlib tests for these behaviors and documents the change in the changelog.

Changes:

• Add a new comma-splice classifier (comma_splice_watch / comma_splice_resumptive) plus a comma_splice_rate report, and upgrade prose scanning to be fence/directive-aware with paragraph-start/end tracking.
• Extend ml_repair.py to apply the new splice repair using ml_metrics.split_resumptive_splices.
• Add fixture-based unittest coverage and document the update in CHANGELOG.md.

File	Description
experiments/​ml-benchmark/​scripts/​test_ml_lints.py	New unittest fixtures covering round-4 splice repair and updated prose scanning/terminal punctuation behavior.
experiments/​ml-benchmark/​scripts/​ml_repair.py	Adds a 4th deterministic repair (comma-splice split) and wires it to the lint classifier output.
experiments/​ml-benchmark/​scripts/​ml_metrics.py	Adds splice detection + rate reporting; rewrites prose scanning to properly handle labels, directive bodies, and CommonMark-style fence closing.
CHANGELOG.md	Documents the new Malayalam benchmark tooling behavior under [Unreleased].

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.