From da3a4f1b4d1cfb02fb1e414609747e71ef50d889 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Wed, 11 Nov 2026 21:13:01 +0800 Subject: [PATCH] docs(WO17): improvement backlog - grammar single-sourcing, ruby parity closure, additivity in maps CI, WF-1 adoption sweep, conformance depth --- TODO.all-sys/17-improvement-backlog.md | 73 ++++++++++++++++++++++++++ TODO.all-sys/INDEX.md | 4 ++ 2 files changed, 77 insertions(+) create mode 100644 TODO.all-sys/17-improvement-backlog.md diff --git a/TODO.all-sys/17-improvement-backlog.md b/TODO.all-sys/17-improvement-backlog.md new file mode 100644 index 0000000..6e715cf --- /dev/null +++ b/TODO.all-sys/17-improvement-backlog.md @@ -0,0 +1,73 @@ +# WO17 — improvement backlog (post ISC-WF-1 + ALA-LC batches) + +Status: BACKLOG — ranked, evidence-linked; the next /goal executes it. +Everything else in this ledger is landed or owner-held. + +## 1. Grammar single-sourcing (parsanol v2 → TS/Py codegen) — structural + +The four runtimes hand-sync their grammar copies (ruby + rs vendor +`isc.parg`; ts + py are hand ports). The drift class is proven, not +hypothetical: TS shipped override rows as bare stage rules until the +maps pilot exposed it (§4.2.6 amendment + ts test); rs keeps +`ruleset` out of its keyword table (documented §2 divergence). +Direction: `.parg` stays the single source; the TS and Py parsers are +GENERATED from it. Land incrementally: (a) cross-grammar parity +harness (parse the 428-file corpus + WF-1 fixtures through ruby and +the candidate generator, require structural IR equality); (b) switch +the TS parser to generated output behind full-corpus + conformance +parity; (c) same for Py. The rs vendored copy joins when its grammar +hosting moves to parsanol-rs's published artifact. + +## 2. Ruby parity closure — bug class + +`Item::Alias#max_length` raises NotImplementedError; 33 golden rows +diverge across 4 Arabic maps (un-ara-Arab-Latn-1971/1972, odni-rus +Arabic variants — the ruby gate lists them as pre-existing). Fix +max_length for alias items in the reference runtime; the +cross-runtime comparison drops to zero deltas outside nothing — the +standing 33-row exemption retires. Prove with +`ruby -Ilib tools/isc_wf1_golden_gate.rb --parity` reaching 0 and the +rs golden unchanged. + +## 3. Additivity as infrastructure (maps CI, all four runtimes) + +The golden harness lives in rs tests (sibling-gated) with ad-hoc +mirrors: ts `test/golden-isc-wf0.test.ts`, ruby +`tools/isc_wf1_golden_gate.rb`. py has NO permanent test (verified +manually during #833). Promote: a maps-CI job that, on every PR, +runs all four gates against `tests/golden/isc-wf0.jsonl` +(12,337 rows / 428 maps) + a permanent golden test in interscript-py +mirroring ts's (sibling-gated, adopting-maps exempt list honored). +Also cover the rs quirk while there: `iscrs capture` skips the +ruleset-only `thai-latn` lib at direct load (no stages) — the loader +should accept ruleset-only docs. + +## 4. WF-1 adoption sweep (reuse the new language) + +Candidates with proven duplication: (a) the Arabic-script +sukun-strip / mark-order / long-vowel machinery now copied across the +11 batch-2 maps (din-ara technique notes) — shared ruleset + per-map +overrides, byte-identical per-pass proof required; (b) ODNI/UN +Cyrillic table overlaps; (c) Devanagari medial tables (partial share +— overrides heavy, measure before committing). W3's first real +adoption: only with document evidence (the ISO 9 pilot dissolved on +PDF evidence; ICAO 9303's split multi-language map #168 is the +candidate to re-examine). Each conversion: adopting-maps.txt entry + +golden zero delta (refactors) or justified deltas (semantic, owner- +visible). + +## 5. Conformance depth + +Embedded tests cover rows the documents exemplify. Where the source +document has a machine-readable text layer, generate exhaustive +per-row tests (every table row, upper+lower) — batch-3's chu/rue did +this by hand (176 rows); make it a tool. Test-from-source stays the +rule; generated tests get a generation-commit note. + +## Owner-held (unchanged) + +PyPI publisher registration + first py version (workflow in place, +ruby #839); Hebrew vocalization data; urd-g2p license_review; B3 +stage-2 ship (MIXED verdict); ISO 259/7098/15919 source acquisition; +release-assets pipeline retire-or-revive (#429); monorepo vestigial +disposition. diff --git a/TODO.all-sys/INDEX.md b/TODO.all-sys/INDEX.md index 695350d..6f46bf4 100644 --- a/TODO.all-sys/INDEX.md +++ b/TODO.all-sys/INDEX.md @@ -61,6 +61,10 @@ versions are owner decisions. DECISION (urd-g2p license_review gate) - [ ] 12-owner-gates.md — Hebrew data, PyPI, urd-g2p license, corpus duplicates (record; owner decides) +- [ ] 17-improvement-backlog.md — the post-WF-1 improvement round + (grammar single-sourcing, ruby parity closure, additivity in + maps CI, WF-1 adoption sweep, conformance depth) — ranked, + evidence-linked, next /goal executes it ## Corpus quality state (2026-11-09)