Rust rewrite: full parity port of overlay (Linux GTK reader + portable core/FFI) - #1
Merged
Conversation
…test Ground-up port of the overlay study logic into a pure Rust workspace: - crates/core: canon (66 books, frozen tok stamp), VRef + refKey, the tokenized KJV corpus (JSONL loader + canonical-order validation + chapter index), Strong's dict + occurrence index + proper-noun heuristic, four-tier search (exact/variant/lemma/typo + phrase + ref + bare-Strong's), and the weave graph (canonical links, BFS components, union-merge, v2 JSON). 28 tests. - crates/layout: reader layout + per-word hit-testing as a platform- agnostic algorithm over injected text measurement (Measure trait), so hit regions always match what the native UI paints. 3 tests. - crates/rnd: feature-gated R&D capability flags (off by default). - crates/ffi: C ABI stub (version probe) for future UniFFI/csbindgen. - apps/desktop: headless core->layout smoke test (GTK next). Verified against overlay's real data: 31,102 verses load, John 3 lays out to 801 boxes, John 3:16 per-word Strong's resolve. Patches/rules dropped by decision. Refs PLAN.md. 33 tests pass, no warnings.
The first native UI (decision: GTK4 on Linux first). A thin shell over pure-core + pure-layout: - Header nav: book dropdown, chapter spinner, prev/next, live title. - Scripture canvas measures + paints with one cairo context (so per-word hit regions match the glyphs exactly), driving pure-layout's display list: gold verse numbers, gold underline on Strong's-tagged words, italic-gray KJV-supplied words, ¶ paragraph breaks. - Click a word -> pure-layout hit-test -> Strong's entry (lemma, translit, pronunciation, def, KJV renderings) in the study side panel. - gtk4 0.11 / libadwaita 0.9; warm-paper theme. Verified visually against overlay's real 31,102-verse data (John 3). PROGRESS.md records the landing point and next steps (Windows/WinUI + Android/Compose shells, both pending a one-time rustup + NDK/mingw install). 33 tests pass, no warnings.
…B Garamond Three arcs on the rust-rewrite branch. pure-ffi — the one C ABI both native shells consume (was a stub): - opaque PureEngine/PureDisplayList handles, callback-based Measure (layout stays in Rust, measured by the platform), stable camelCase-JSON payloads, panic-safe (catch_unwind) boundaries, explicit ownership/free. - open / open_from_bytes, toc, chapter_count, verse/token, layout_chapter, hit_test, strongs (+ occurrences), search. - generated bindings: C header (cbindgen) + C# P/Invoke (csbindgen), plus a hand-written idiomatic C# wrapper and a Kotlin/JNA wrapper. Regenerate with `cargo run -p pure-ffi --features bindgen`. - verified: 10 Rust ABI tests, a C consumer (bindings/c/smoke.c) and a C# P/Invoke consumer, both driving the real 31,102-verse corpus. - reviewed adversarially (6 lenses); fixed the surfaced wire-contract issues (verse key, exported flag constants, callback totality + NaN clamp, null on out-of-range chapter, explicit-null item fields, u32->u16 truncation). - Windows x86_64 .dll cross-builds (mingw); host .so/.a build. core::notes — the kjv-notes.jsonl loader (missing file -> empty, bad lines skipped, notes grouped per verse in file order), feeding search + the reader. desktop reader — the core's study surface, wired into GTK: - multi-tier search box; concordance (occurrences); 1769 margin notes; Ctrl zoom; keyboard scroll + chapter nav. - real typography: Pango shaping in the bundled EB Garamond (registered with fontconfig at startup, no system install), replacing cairo's toy font API. - fixes: hit regions align with glyphs (was offset by the top margin -> hit the line below); Strong's on double-click (single click reserved); study panel is on-demand (hidden by default, Esc/clear-search closes it). cargo test: 44 green across core/layout/rnd/ffi, no warnings.
Load all weaves from OVERLAY_HOME/weaves (+ suggested) at startup and precompute a verse → partners map (both directions of each undirected link, deduped). - Reader: verses carrying cross-references get a subtle gold gutter dot. - Study panel: double-clicking a word now lists that verse's cross-references — each a clickable jump link, tagged with the weave that asserts it — alongside its Strong's entry, margin notes, and concordance. Reuses core's ported weave graph + loader; the authoring / constellation / chords UI remain later work.
Refactor the single-pane reader into N independently-navigable columns (overlay's core reading model). - Each pane has its own book/chapter nav strip (dropdown, chapter spin, prev/next), scroll position, draw, and per-word hit-test. - Per-pane + / ✕ add and close columns (up to 3); the pane row rebuilds on change so handlers always hold a valid index. - The active pane (last touched; gold top accent) is what the header search, cross-reference/search "go" links, and the study panel target. - Ctrl+scroll / Ctrl+/- zoom all panes together; Page/Home/End scroll the focused pane; ←/→ (or [ ]) step its chapter. Verified: John 3 + Numbers 21 render side by side; single-pane nav strip shows book/chapter/prev/next/+ . cargo test 44 green, no warnings.
Overlay a transparent, click-through draw layer on the pane row and draw a soft
gold curve between the two endpoints of every weave link whose verses are both
currently on screen. Verse positions map into the overlay via GTK
`compute_point`, so the connectors track each pane's scroll automatically; the
layer redraws on scroll, navigation, zoom, and pane add/close.
Endpoint + gutter dots mark linked verses. Verified with Genesis 15 ∥ Romans 4:
a curve joins Gen 15:6 ↔ Rom 4:3 ("Abraham believed God … counted … for
righteousness"). cargo test 44 green, no warnings.
- Hover a Strong's-tagged word for a quick tooltip gloss (code + lemma + transliteration + trimmed definition), wired through GTK's query-tooltip and the same per-pane hit-test as double-click. - A canon-overview strip under the panes maps the 66 books in their 8 canon sections (Law … Revelation) with the OT/NT divide and a pin per pane (active in gold); clicking a book jumps the active pane there. Verified: strip shows the sections + John's pin under Gospels. cargo test 44 green, no warnings.
core::thread and core::tag — port overlay's Thread.hs / Tag.hs (personal study
data, one JSON file per item under home/threads and home/tags):
- Thread { name, notes, entries:[ref, span, word snapshot, note, added] };
Tag { name, color?, members:[target(verse|concept), note?, added] }.
- Format-checked, ref-parsing serde (custom (de)serialize keeps the frozen
"overlay-thread-v1" / "overlay-tag-v1" wire form); dir loaders sorted by name,
bad files reported not dropped; membership queries. +7 tests.
Desktop: "Threads" / "Tags" header buttons open a browse list in the study
panel; a thread shows its passages (snapshot preview + note) as jump links, a
tag shows its verses (jump) and concepts (concordance); a verse's tags also
appear in its word study. Verified against the Romans Road thread.
Read side only — authoring (add/tag/write) is next. cargo test 51 green.
core::store — a portable atomic writer: write to a sibling temp file, fsync,
close the handle (required on Windows), then fs::rename over the target
(replaces the destination on Unix and Windows alike). All paths built with
Path::join; a `slug` helper mirrors overlay's filename slugging.
core authoring — thread::{thread_file, write_thread, add_to_thread} and
tag::{tag_file, write_tag, add_member, remove_member}: create-or-append by
case-insensitive name, dedupe tag members, refuse to clobber an unparsed file,
caller-supplied timestamps (keeps core deterministic). +4 tests covering the
write→reload round-trip in a temp dir.
Desktop — a word study now offers "＋ tag verse" and "＋ add to thread" (a modal
name prompt, then create-or-append, atomic write, and reload) and "✕" to untag;
the data home is kept in State for reloads. Verified the action links render;
core write path is unit-tested end-to-end.
cargo test 55 green, no warnings.
core::weave — write_weave + add_link (find-or-create a weave by case-insensitive name, dedupe/canonicalize via Weave::add_links, refuse to clobber an unparsed file, caller-supplied timestamp), through the atomic cross-platform store. +1 test covering create → append → dedupe round-trip. Desktop — single-click a word to pin its verse as a link endpoint (blue band); when two panes are pinned the header "＋ link" button enables. It prompts for a weave name, writes the whole-verse link, reloads weaves + rebuilds the cross-ref index and connector pairs, and the new gold connector line draws immediately. Double-click still opens Strong's; the loaded weaves are kept in State for find-or-create. Read + write of the weave graph are now both in the reader. cargo test 56 green.
The C ABI was read-only; the Windows/Android shells could display study data but
not create it. Now the engine loads notes + threads + tags + weaves from its
home dir (when opened via pure_engine_open) and exposes:
read — pure_engine_threads_json / _tags_json / _verse_xrefs_json(ref)
write — pure_engine_thread_add / _tag_add / _tag_remove / _weave_add_link
(null on success, else an owned error string; caller supplies the UTC
timestamp; writes go through core::store's cross-platform atomic write;
the engine reloads its in-memory copies after each write)
An engine opened from bytes has no home and returns an error from the write
calls (study data stays read-only). A dedicated guard turns any panic in a write
into an error rather than a false success. New wire DTOs for threads/tags/xrefs;
regenerated the C header + C# P/Invoke shim.
Verified with an end-to-end ABI test: open a temp home, tag/thread/weave via the
ABI, read back, exercise the error paths. cargo test 57 green; Windows x86_64
.dll still cross-builds.
core::weave gains is_suggested / approve_weave / reject_weave: approve marks every link approved and lands the weave in home/weaves (merging a same-named weave there), removing the suggestion; reject deletes the file. Both go through the cross-platform core::store atomic writer. GTK: a "Suggested" header button opens the review queue — each proposal under weaves/suggested with its links as jump links and approve/reject actions that reload + re-render. FFI: pure_engine_suggested_weaves_json lists the queue with per-item ordinals; pure_engine_weave_approve / _weave_reject act by that ordinal (null on success, else an owned error; needs a home). Regenerated the C header + C# P/Invoke shim. 60 tests green (incl. an end-to-end ABI review round-trip); Windows x86_64 .dll still cross-builds.
…/spanB A click now pins a single word as a link endpoint; clicking another word in the same verse grows the span from the first click's anchor (widen or narrow without clearing). The pinned span highlights its exact words (blue) rather than the whole verse, and make_link records each endpoint's token span via Link::canon_span — the core already stored spanA/spanB. Connectors and cross-references stay verse-level, so nothing else shifts.
core: thread::set_thread_notes / set_entry_note (empty clears; errors on missing thread or out-of-range entry) and weave::set_weave_notes (marks notesSource=hand). All via the cross-platform atomic writer. GTK: a thread view gains "✎ notes" and a per-entry "✎ note"; a suggested weave gains "✎ note". A new prompt_text dialog pre-fills the current text and permits an empty submission (to clear). FFI: pure_engine_thread_set_notes / _thread_entry_set_note / _weave_set_notes (null on success, else an owned error; needs a home), covered by the ABI round-trip test. Regenerated C header + C# shim. 62 tests green; Windows .dll still cross-builds.
weave_markup renders a weave's kind, notes (with ✎ note), and each link as its two passages with the spanA/spanB words bolded (KJV-supplied words stay italic-gray, as in the reader) and the shared label shown. Opened from a cross-reference (the weave name is now a link) or the suggested review list (a new "⇔ compare" action). Read-only view over the corpus; no new on-disk format.
navigate_pane cleared the pane's display list and now polls every 8ms (up to ~1s) until the new chapter is painted and its scroll extent has grown to the new content height, then scrolls once. This fires as soon as the layout is ready — usually the next frame — instead of hoping a fixed 50ms delay lands after the paint, and can't scroll a stale chapter or clamp short against a not-yet-updated adjustment.
core::config: a cross-platform per-user config (Windows %APPDATA%, macOS Application Support, else XDG) holding the study mode + body size, read and written through one code path (atomic write via core::store). Loads default-as-first-run when absent; a damaged file loads default without re-prompting. GTK: first launch shows a guided chooser — Simple reader vs Full study. Simple hides the study/authoring surface (Threads/Tags/Suggested/＋link buttons, the verse's tag+author actions, weave-note editing), leaving a clean reader + search + Strong's; Full unlocks it. A toolbar toggle flips modes; the choice and the live zoom size persist across launches. 65 tests green; Windows .dll still cross-builds.
core::home resolves the data directory in order: $PURE_STUDY_HOME / $OVERLAY_HOME → the CWD if it's a working tree (has data/kjv.jsonl) → next to the executable → the per-user data dir (XDG / %APPDATA% / macOS Application Support). An explicit env override is honored verbatim so a precise "hydrate here" error can surface; otherwise only a directory that really looks like a home is used. The desktop app uses it (no more hard OVERLAY_HOME requirement — verified launching from a checkout with the env unset), logs the resolved home + kind at startup, and its data-missing screen now lists every location searched. Paths join cross-platform throughout.
A GitHub Actions workflow with three jobs: - test: cargo test the portable crates (core/layout/rnd/ffi; the GTK shell needs system libs and is excluded). - bindings: regenerate the C header + C# P/Invoke shim and fail on any diff — the doc-vs-wire drift class the ABI review flagged. - windows-dll: cross-build the C ABI for x86_64-pc-windows-gnu (mingw), matching the local cross-build. Verified the test and binding-drift steps pass locally (bindings in sync).
Two R&D capabilities that need no ML packs — both derived from data already loaded — land behind the Full-study gate: - core::crossref: parse the hydrated TSK topical cross-reference tier (data/cross-references.tsv, ~344k vote-ranked refs, ranges supported; missing file → empty, the reader runs fine). Shown labelled "study cross-references (TSK)" in the word study. - rnd::bridge (feature `bridge`): the OT↔NT etymology bridge — Strong's own "of Hebrew origin (Hxxxx)" derivations linking Greek↔Hebrew codes. Surfaced as "↔ cross-testament" partners (concordance links) per word. Wired into the desktop (Full study only) and verified against the real corpus: Matthew 1:2 "Abraham" shows G11 ↔ H85 and 23 TSK cross-references with verse ranges. CI now also runs the bridge-feature tests. Docs + module map updated. The heavier R&D (concept engine, embeddings, morphology, keyness, trust, quotation detection) genuinely depends on the ML data-pack pipeline and stays deferred. 72 tests green (70 default + 2 bridge).
rnd::embed (feature `embeddings`) loads the offline-trained concept-vectors.vec (word2vec text format) with its .meta (tokenization gate + aligned flag + alias map) and .freq sidecars — no training, pure load. Rows are unit-normalised so cosine is a dot product; alias keys resolve to their root's row (the split Strong's numbers the KJV tags). Queries: nearest_concepts (same-language distributional near-synonyms) and cross_concepts (the cross-testament semantic bridge, live only when the space is Procrustes-aligned). parse_embedding is split out for file-free unit tests; verified against the real 7426-vector artifact. Surfaced in the GTK word study (Full study only): "≈ concepts near" and "≈ across the testaments", each a concordance link with its lemma — e.g. G26 ἀγάπη → διακονία/ὑπομονή/γνῶσις and Hebrew neighbours. Absent artifact → the section simply doesn't render. Tests: 6 with --features "bridge embeddings"; no warnings.
rnd::morph (feature `morphology`) ports the consuming side of overlay's Morph.hs: the OSHM (Hebrew/Aramaic) and Robinson (Greek) parsing-code parsers, the study-panel renderer, and the sidecar loader over morphology.jsonl. The offline projection that builds the sidecar stays in Python; this only reads it. Codes parse once at load; a stale tokenization stamp is refused; a missing file → None. Surfaced in the GTK word study (Full study): the token's original-language parse under the English word — e.g. Genesis 1:1 "created" → "Qal perfect, 3rd masculine singular", alongside its concept neighbours and the cross-testament partner G2936 κτίζω. Tests: 11 with --features "bridge embeddings morphology" (5 new: Hebrew/ Aramaic/Greek render, error-not-panic, sidecar gloss). CI runs all three.
VerseSim ports the Arora-Liang-Ma SIF recipe over the concept embedding: weight each concept by a/(a+p(concept)) with the trainer's own frequency table, average per verse, subtract the per-testament mean, then remove the projection onto that testament's top principal component (power iteration, no linalg dependency; skipped below 50 verses). similar_verses_in ranks the same testament by cosine; similar_verses_cross the other (aligned only). Built once at startup from the embedding + corpus, surfaced in the GTK word study (Full study) as "verses like this" — e.g. John 3:16 → John 6:27, 3:17, 3:15, 12:25. Unit test covers the power-iteration PCA. 12 tests with the R&D features on.
data-prep/README.md records exactly which artifacts the R&D tier consumes (concept-vectors.vec/.meta/.freq, morphology.jsonl, cross-references.tsv), how each is produced by the existing offline Python (overlay ml/ + pipelines/), and the key fact that it is build-once (no GPU, no foundation model, one lightweight CPU skip-gram) → ship-the-files. Notes licensing (concept vectors wholly owned; TSK CC-Attribution; OSHB CC-BY-4.0; domains excluded for copyleft) and that hosting the pack is just placing the files under <home>/data/, which core::home already resolves. PLAN/PROGRESS module map + status updated for rnd::embed, rnd::morph, core::crossref, and the documented pack.
New crates/hydrate CLI. `copy --from <src> --to <home>` places the pack (core kjv/strongs/notes + R&D cross-references/concept-vectors[.meta/.freq]/ morphology) into <home>/data; `check [--home]` verifies each artifact by loading it through the real code paths and reports what will light up (verse/entry counts, TSK coverage, embedding dim/alignment/freq, morphology coverage, etymology-bridge code count), exiting non-zero only when a required core file is missing. It places + validates, it does not train/generate (that's the offline pipeline; see data-prep/README.md). Verified: copy from the overlay checkout hydrates a fresh home to 31102 verses, 343556 TSK refs, 7426 aligned vectors, 31091 morph verses. Added to the workspace + CI test set.
… pack rnd::bridge gains the fused layer over the etymology base: load_sources reads the committed bridge/*.json witnesses (LXX / Abbott-Smith / TIPNR) plus optional hydrated data/bridge-sources.json + quotation-pairs.json; Priors loads source-priors.json. FusedBridge::build merges etymology + external partners per lemma, each tagged with the witnesses that assert it and ranked by best trust prior (etymology treated as authoritative). pure-hydrate now copies the bridge sources + priors and reports the fused link count. The desktop uses FusedBridge and shows cross-testament partners with provenance — e.g. H4899 מָשִׁיחַ → G5547 (abbott-smith+lxx) = Christ, G3323 (etymology). Verified against the real pack (5922 external links). 13 rnd tests with all features (3 bridge, incl. the fusion merge/ranking).
The engine loads the R&D artifacts at open (fused bridge from strongs + home witnesses; concept embedding + morphology sidecars when a home is present) and gains four read endpoints: - pure_engine_concept_neighbours_json(code,k) — near + cross-testament - pure_engine_bridge_partners_json(code) — fused partners w/ sources+prior - pure_engine_morph_json(ref,tok) — parse code + rendered gloss - pure_engine_similar_verses_json(ref,k) — SIF, model built lazily + cached Each returns null when its artifact is absent (bytes-opened engines get etymology-only bridge, no embedding/morph). New wire DTOs; regenerated the C header + C# shim. pure-rnd is pure Rust so the cdylib still cross-builds for Windows (verified). End-to-end ABI test with tiny artifacts covers all four + the null paths. 13 FFI tests. Now the Windows/Android shells reach the full study + R&D surface over the one ABI, matching decision #1.
Two reader chrome bugs: - The Paned divider was a fixed absolute 700px from the left, so on a wide window the study panel (end child) ballooned to fill the rest. Now the panel opens as a stable ~380px sidebar: on the hidden→visible transition show_study sets the divider to width − PANEL_WIDTH. - The chrome (header nav, dropdowns, study panel, canon strip) followed the dark system theme, leaving light text on the light nav strip (illegible). The reader is a warm-paper light design, so force the light color scheme (adw StyleManager ForceLight) to match the cream scripture + gold accents. Verified: panel opens as a sidebar (not massive) and the nav is legible dark-on-cream.
Ports the reader-facing core of overlay's Concept.hs (feature `concept`, pure corpus folds, no ML data): per-concept occurrence stats + book distribution (top_books / testament_split), and the collocation graph — co-occurrence → PPMI → mutual-kNN → label-propagation communities. Memory is bounded by keeping only the mutual-kNN edges (the dense PPMI matrix is dropped after filtering). The Concept engine exposes stat / top_books / testament_split / collocates / community, which back the concept neighbourhood diagram and the dispersion strip (viz next). Also documents the config-window Hyprland rule work in this session: pure-study now opens on the little Samsung (workspace 1 → HDMI-A-2). 17 rnd tests with all features (4 new: distribution, ppmi, mutual-kNN, communities). CI feature set updated.
…ion) The desktop builds rnd::concept at startup (feature `concept`) and, in Full study, the word study now shows two more sections per Strong's code: - ◦ collocation field: the code's co-occurrence community (what shares its verses), distinct from the distributional embedding neighbours. - ◦ distribution: the books it concentrates in (top 5) + the OT/NT split. Verified against the real corpus: H430 (Elohim) → Deuteronomy/Psalms/ Genesis with its Hebrew collocation field. (The radial graph + cairo dispersion-strip visualizations remain as polish; the data is surfaced.)
rnd::burst ports Burst.hs: a Poisson scan statistic over each concept's occurrence positions (regularized incomplete gamma / Lanczos gammaln, Numerical Recipes §6), Bonferroni-corrected within a concept and Benjamini–Hochberg–gated across concepts, to surface deliberately-packed theme-words. Pure corpus statistics. The word study now shows a "leitwort" line (span + tightness) when the clicked concept bursts. To keep launch instant, the heavy analytics — SIF verse-sim, the concept graph, and the leitwort sweep — are no longer built eagerly at startup; ensure_analytics builds them once on the first Full-study word lookup and caches them (a no-op in Simple mode). Time-to-window dropped 10.6s → 3.3s (debug); the first deep lookup pays the one-time sweep. 10 rnd tests with the concept feature (3 new: gamma/Poisson, scan window, Benjamini–Hochberg).
core::config gains the reading session — the open panes (book, chapter) and the active index — folded into the same cross-platform config file (additive fields, older configs still load). The desktop restores those panes at launch (falling back to John 3 on a fresh install, active index clamped into range) and saves them on window close-request via persist_config (which now snapshots panes + mode + body size). Verified: a config naming Romans 8 + Psalms 23 (active 1) reopens both panes with Psalms 23 active.
A "Map" study-tool (Full mode) opens a popup arc diagram over the canon- ordered axis: each woven book pair is a ribbon whose thickness is the link count and colour marks OT-internal (gold), NT-internal (blue), or cross- testament (purple). chord_arcs aggregates State.links by book-order pair; draw_chord_map paints section bands + the OT/NT seam + heaviest-first ribbons; click a book to jump the active pane, Esc closes. Verified against the real weave library (cross-testament arcs dominate).
…xCache.hs) Two startup wins: - core::corpus caches the parsed corpus beside the source as a gzipped bincode snapshot (kjv.jsonl.idxcache), keyed by the source's length + mtime + tokenization; any mismatch re-parses. Skips the ~19 MB JSONL parse on unchanged data. store gains write_atomic_bytes for it. - The desktop no longer parses the multi-MB embedding + morphology at launch — they load lazily in ensure_analytics with the other Full-study analytics, so a Simple reader never pays for them. Time-to-window (debug) 3.3s → 2.7s warm; the cache is 6 MB gzipped (was 37 MB raw). Corpus cache round-trip/invalidation covered by a test.
rnd::witness (feature `bridge`) loads data/text-witness.json — a lexicon-free H↔G alignment that can disbelieve bridge links — and enforces the gate: disbelief() returns a percentile ONLY when the witness has passed its held-out grading (witnessQualified && testimonyActionable), so an ungraded witness structurally cannot accuse. Missing/corrupt file → silent. The desktop loads it and marks a disbelieved cross-testament partner with a ⚠ (heb-first keying). The shipped witness is not yet qualified, so nothing surfaces today — the path is complete for when a graded one ships. 23 rnd tests with all features (3 new witness: silent-when-unqualified, strongest-percentile-when-qualified, missing-file).
A "▸ concept map" link in the Full-study word study opens a popup for the clicked code (ConceptGraph.hs + ConceptMap.hs): a radial neighbourhood with the concept at centre and labelled spokes to its distributional embedding neighbours (gold) and its collocation community (green), over a dispersion strip shading the 66 books by density. Esc closes. Verified with G26 ἀγάπη.
Word study double-click surfaced R&D concept lists by bare Strong's
code + original-language lemma only — illegible to an English reader,
and the six analytic tiers ran together as one undifferentiated block.
- english_gloss(): learn the recognisable English word from how the KJV
actually renders a code (mode over sampled occurrences, translator-
supplied words skipped), not Strong's etymological headword. Falls
back to distilling the definition. Deterministic across panel + map.
- concept_links(): render English-first ('love ἀγάπη'), code as the link.
- Sectioned analytics: SAME ROOT ACROSS TESTAMENTS / SIMILAR CONCEPTS /
APPEARS ALONGSIDE / WHERE IT CONCENTRATES / LEITWORT, each a spaced
header with blank-line separation; near + cross concepts merged.
- Distribution shows counts ('Genesis ×12'), not chapter-like refs.
- Leitwort reworded ('42 of its 55 uses cluster in …', p ≈ 10⁻ⁿ) instead
of a bare 'strength' number; witness flag worded 'disputed by usage'.
- Bridge sources humanised (lxx → Septuagint).
- Radial concept-map node labels English-first to match.
- Test: panel markup shapes validated through Pango's parser; distil_gloss.
The C ABI now cross-builds for ARM64 Windows in CI, not just x86_64 — the target a Snapdragon/Surface Windows box uses natively. README documents the portable-vs-GTK split and the exact pull-and-build steps on ARM Windows.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adopts the Rust rewrite as
main: the pure domain core, layout engine,feature-gated R&D tier, the single C ABI, the hydrate CLI, and the GTK4/libadwaita
Linux reader — at feature parity with the overlay backlog.
Highlights
core/layout/rnd/ffi/hydrate): dependency-light pureRust, cross-platform file I/O (atomic temp→fsync→rename, per-OS dirs).
similarity, morphology, keyness/leitwort (Poisson scan + BH-FDR), text-witness.
session restore, gzipped corpus cache (0.63s warm release launch).
from actual KJV renderings, sectioned analytics with headers, humanised wording.
aarch64-pc-windows-msvcin CI.Verification
aarch64-pc-windows-msvclocally; CI links theARM64 + x86_64 Windows
.dll.🤖 Generated with Claude Code