ci: send the commit message and a data summary to wikitermbase

[forzagreen]

The pull request that wikitermbase opens after a dump change (e.g. forzagreen/wikitermbase#110) only carries the arabterm commit SHA, so it does not say what it is for. This sends it what it needs; forzagreen/wikitermbase#111 is the receiving side.

What changes

• arabterm/scripts/summarize_changes.py (new, stdlib-only like validate_db.py): compares two versions of arabterm.db and writes a Markdown summary — dictionaries added / removed / changed, terms added / removed / modified per dictionary, the columns that were modified, and the first 20 terms of each kind. Timestamps and nbr_entries are ignored. A term moved to another dictionary counts as removed from the first and added to the second.
• notify-wikitermbase.yml: unpacks arabterm.db.gz as it was before the push and as it is now, runs the script, and adds message (the head commit message) and summary to the repository_dispatch payload next to arabterm_sha. The summary is also written to the job summary.
• CLAUDE.md and README.md describe it.

The summary step is continue-on-error: if it fails, the dispatch is still sent, without a summary.

Example

For #58 the summary is:

Data changes

• Dictionaries: 85 → 86 (+1)
• Terms: 509871 → 509973 (+102)

Dictionaries added (1)

• aala_interdisciplinary_linguistics · Dictionary of Interdisciplinary Terms: Linguistics (AALA, 2026) · معجم المصطلحات البينية: مجال اللسانيات (المجمع الجزائري، 2026) · Q141645945 · 102 terms

and for #56 it names the dictionary and shows the term: arabic: تسمم ةكحو لي → تسمم كحولي.

Notes

• The squash-merge message of a data PR becomes the title and description of the wikitermbase PR.
• The summary describes this repository's push. If an earlier refresh PR is still open in wikitermbase, the diff of the new one also contains that earlier change.
• Either side can be merged first: the current wikitermbase workflow ignores the extra payload fields.

Testing

The script was run on four past changes: #58 (new dictionary), #56 (one term fixed), #51 (page numbers of energy_terms) and #50 (the linguistics split). The dispatch step was run locally with gh stubbed, to check the payload. The workflow itself has not run on GitHub Actions: it only fires on a push to main that changes the MariaDB dump, so the first real run is the next data merge.