lang-smn

Speller: keep a->o and á->i at the default weight giella-core 1.16.12 fixed editdist.py so a pair listed once also sets its reverse. The reverses a for o and á for i lose 2 top-1 rows and 1 top-5 row (kannâd-uv -> kannat-uv loses to konnâd-uv); listed at 10 they cost what they did. typos-default-generated.tsv (1,002 rows): 854 / 947 / 952, as before the fix.

Rebuild with the lang-deps step (divvun-actions#25)

sed shared lexc dirs in local mods

Trigger build: gzipped speller accuracy, optimized accuracy viewer

Trigger build: gzipped speller accuracy, optimized accuracy viewer

Rebuild to publish pkg-version.json

Trigger rebuild

Update list of (non)passing tests

Apply changes 343

Retrigger docs-data publish after giella-core dependency fix

Trigger build: fst-version.json -> version.json

Trigger build: fst-version.json -> version.json

Enable boundary edits at weight 20 The only runtime knob in a full config sweep that moves reachability rather than reordering: it recovers missing-space and missing-hyphen rows (ovdilmainâšum -> ovdil mainâšum, suomâugrâlii -> suomâ-ugrâlii) that no letter edit reaches. Plateau 10-30, zero row regressions on either test set, +3 combined and +8 generated top-1, about +5% average lookup latency. word-split-weight reaches the same rows with worse ranks and adds nothing on top, so boundary-edit alone. Also swept and left alone: beam is inert from 20 to unbounded (top-1/top-5 identical at every value), and the 3/2/3 reweight stands - the nominally better 5/0/3 costs the generated set 22 top-1, a classic acceptance-set overfit. This file ships inside the archive.

Set maxweight to 20 for the canonical speller corpus Fitted against the canonical corpus - corpus-smn plus corpus-smn-x-closed, 4,221,192 tokens assembled at build time - which requires giella-core 1.15.2 or later; before that, a plain build silently weighted from the in-tree file and this value would be measured against the wrong corpus. Note smn is the one language whose PUBLIC corpus alone is smaller than the in-tree file, so a build without the closed corpus inverts every comparison here. Swept 5..40 with unit refinement across 16..22. Versus the previous value of 10, on the same canonical corpus: typos.tsv +12 top-1 / +4 top-5, generated unchanged top-1 / +2 top-5. 20 is the centre of an 18..22 plateau, the peak of the even blend, and effectively tied for the peak under attested-weighted blending; it is also best top-5 on all three test sets at once. Separately from the tuning: the canonical corpus itself cost the generated set 53 top-1 hits relative to the old in-tree corpus, and no maxweight value recovers them - a corpus effect, not a weighting effect. Ranking is close to exhausted on both sets (top-5 within a point of its reachability ceiling); the remaining headroom is reachability.

Add more real typos, extracted from the corpus files based on potential typos (Err/Orth entries) - these typos are actual ones found in corpus material

Clean README and trigger rebuild

Trigger build to generate new gramcheck docs data and fix private repo badges

Merge branch 'smn-speller-tuning'

fix fallback lexc generation for smi

Trigger build to test buildkite-overview badge fixes