feat(leaderboard): add (cov.) scoring-scope toggle

[malteos]

Summary

• Add a radio between the dataset metadata and the table per tab. Default = paper-headline view; (cov.) restricts macro/micro/FPR/Languages to gold samples whose language is in the model's declared support set, mirroring the CommonLID paper's (cov.) column.
• Persist supported_languages in summary.json (schema v3): sorted ISO 639-3 list when the model can enumerate, JSON null for LLM-style models whose support set is undefined.
• New standalone script scripts/backfill_supported_languages.py walks existing summary files and writes the field in place (used to populate the HF dataset without re-running any evals — companion PR opens on commoncrawl/commonlid-results).

Cov view re-ranks the leaderboard because models that intentionally cover a subset are no longer penalised for the long tail:

model	macro F1 (all)	macro F1 (cov.)	cov. languages
cld2	49.5	79.3	68
fasttext	49.3	72.6	74
GlotLID	60.4	68.6	96
OpenLID-v2	47.4	68.0	76
cld3	34.3	66.6	56
commonlingua	57.3	66.4	94
funlangid	46.4	58.1	87
pyfranc	39.3	57.2	58
AfroLID	9.2	43.5	23

supported_languages follows the same tri-state convention as LIDModel.discover_supported_languages(): None (undefined / LLM), [] (declared zero), or list[str]. The leaderboard collapses all three "no cov data" cases to em-dashes in the cov view; the persisted distinction between null and missing is preserved on disk.

Coverage counts vs the paper

We compared the supported_languages lengths produced by discover_supported_languages() against both Table 1 (tab:eval_models) and the mutual-coverage diagonal in tab:coverage. Note the paper's two tables disagree with each other for several models, because Table 1 reports the raw upstream label count while the coverage diagonal reports the dedup'd / ISO-conformed count actually used in evaluation.

Model	Paper Table 1	Coverage diag (tab:coverage)	Ours	Δ vs Table 1	Δ vs coverage
GlotLID	1868	1868	1868	0	0
OpenLID-v2	193	193	193	0	0
AfroLID	517	515	516	−1	+1
cld3	99	99	101	+2	+2
fasttext	218	210	210	−8	0
FUN-LangID	1634	1549	1552	−82	+3
CLD2	158	158	172	+14	+14
pyFranc	414	410	379	−35	−31

Per-model diagnosis:

• GlotLID / OpenLID-v2: exact match.
• AfroLID / fasttext / FUN-LangID / cld3: ours equals the coverage diagonal within ±3. Table 1 reports the raw label count (518, 218, 1634, 107 raw → 516 / 210 / 1552 / 101 after _conform()), the coverage diagonal does what we do.
• CLD2 (+14): pycld2 ships 303 raw labels; after dropping xx-Script sentinels (xx-Cyrl, xx-Hani, …) and BCP-47 region variants (sr-ME) we get 172 unique ISO 639-3 codes. The paper's 158 is more conservative — it likely filtered some BCP-47 / script tags whose _conform() output is a real ISO 639-3 code (e.g. zh-Hant → zho).
• pyFranc (−31): upstream pyfranc.franc.data currently exposes 380 unique raw codes; we drop only the 1 duplicate (dan collision). The paper's 410 / 414 reflects an older snapshot of franc — the library shrunk between the paper's submission and now. Not actionable on our side without pinning an older pyfranc.

Recommendation: leave the numbers as-is. discover_supported_languages() is internally consistent (every model goes through the same _conform() filter) and reflects the currently-installed library version — so the cov metric is computed against what each model can actually predict today, not against a paper-table snapshot. Overriding to literally match the paper would mean reporting a coverage set that diverges from what _conform()'d predictions can hit at eval time.

Test plan

• make lint && make typecheck — clean
• make test — 247 passed, 94.56% coverage. New tests cover _row_from_summary cov math, the three "no cov data" cases, the radio change handler, and Result.summary() round-tripping None / [] / list.
• Backfill local data/results/, restart make leaderboard, switch the radio: table + legend swap in lockstep, em-dashes for GPT-* rows, sorted-to-bottom in cov view.
• Follow-up: open the matching PR on commoncrawl/commonlid-results so the deployed Space picks up supported_languages on next restart.