v1.2.0

v1.2.0

A checkpoint-sized batch of corpus expansion, two new language categories, and
the tooling to keep doing it.

data/ grew from 152 repo-folders / 2,344 files to 218 / 3,306. Every added
folder is real, unmodified source copied from a pinned commit of a clone in the
gitgalaxy/data pool, recorded at exact confidence in
tools/independent_data_auditor.py and the per-category SOURCES.md tables — no
files were generated or hand-written. Two categories hit the pool's real-source
ceiling and now say so in their SOURCES.md rather than being padded.

New categories

embedded_python (1 folder / 14 files) — see the python / embedded_python
rows in the table below.

sqlite — 9 folders / 82 files (data/sqlite/SOURCES.md)

GitGalaxy's sqlite language standard is distinct from generic sql (its own
extensions, .sqliterc, sqlite3 shebang, ~50 rules). Real .sql exercising
that dialect: WITH RECURSIVE, STRICT / WITHOUT ROWID, ->> / json_extract
/ json_each, RETURNING, ON CONFLICT DO UPDATE, GENERATED ALWAYS AS, window
functions, FTS5 bm25() / snippet() / MATCH, R-Tree, PRAGMA journal_mode=WAL, sqlite_master, CREATE TRIGGER, ATTACH DATABASE, and the
.-prefixed CLI dot-commands. From sqlite (Public Domain), mediawiki (GPL-2.0),
sqitch (MIT), prisma (Apache-2.0), yii2 / flask (BSD-3), Dancer2.

The pool holds ~80 real SQLite .sql files total — the category is capped there.

_{* v1.1.0 root SOURCES.md said python was 487 files; see the python row.}

Expanded categories

Category	v1.1.0	v1.2.0	What's new
assembly (SOURCES)	3 / 15	16 / 253	NASM preprocessor torture (%macro/%rep, every SSE/AVX/AVX-512/AMX extension), GAS/AT&T (cosmopolitan, xv6, FreeBSD libc + kernel, x86-bare-metal), AArch64 + Darwin ABI + NEON (HelloSilicon), 6502 (cpm65), ARM bare-metal (RaspberryPi — unknown, no upstream license), 16-bit real-mode boot (os-tutorial, linux-1.0), machine-generated perlasm (blst).
groovy (SOURCES)	2 / 9	16 / 329	Was .java-only (Spock's AST transform, Gradle's Java API) — zero files the standard actually classifies. Now real Groovy across its dialects: Gradle build DSL (closure-as-config), @CompileStatic plugin authoring, Spock specs (given:/when:/then:/where:, power asserts, data tables), the Jenkins Groovy view builder-DSL, Jenkinsfile.
cobol (SOURCES)	11 / 308	15 / 589	~2×. The NIST COBOL-85 CCVS positive-conformance suite (150 programs, round-robin across every language module), the che-che4z LSP's own parser fixtures (XML/JSON PARSE, EXEC SQL/EXEC CICS, dialect keywords), Open Mainframe Project's cobol-check, JCICS interop samples, GnuCOBOL library copybooks.
tcl (SOURCES)	1 / 13	9 / 156	The SQLite .test suite — TCL harness (do_test name { db eval { …SQL… } } {…}) wrapping SQL string blocks, the polyglot boundary the standard's "SQLite mega-sensor" targets — across core DML / query planner / CTE-window / JSON-triggers / FTS5 / R-Tree; plus real application Tcl from macports-base (port1.0 DSL, registry2.0).
html (SOURCES)	7 / 19	15 / 67	GLSL-in-<script> polyglot (three.js), deep <section> state trees (reveal.js), framework microsyntax that isn't valid HTML (Angular @if/{{ }}/[x]/(y)), CSP <meta> + nonce templates (VS Code webviews), deliberately-broken DOM fixtures (Playwright), single-file inline-everything apps (Cesium), h5bp source-vs-build.
css (SOURCES)	3 / 9	11 / 49	Tailwind v4 @theme/@layer/@utility custom at-rule DSL, Gutenberg CSS Modules (composes:, :global()), minified single-line Bootstrap, machine-generated RTL sheets, reveal.js compiled themes, vendored third-party CSS, legacy MediaWiki module stylesheets.
python (SOURCES)	6 / 281*	6 / 281	No files changed. v1.1.0's 487 was a bookkeeping error — the issue-#4 audit recorded fastapi at 446 files but that folder has only ever held 240 (git ls-tree v1.1.0 confirms). The v1.2.0 regen made the count honest.
embedded_python (SOURCES)	—	1 / 14	New category. meow_turtle — original RP2040/RP2350 MicroPython firmware: boot.py, import machine, and PIO assembly embedded in Python via @rp2.asm_pio. GitGalaxy's embedded_python standard (.py/.mpy, micropython shebang, machine/UART/viper markers) classifies this separately from CPython. unknown by design.
jcl (SOURCES)	5 / 191	6 / 193	Effectively flat. This category is at the pool's real-licensed-source ceiling — the entire pool holds ~191 real licensed .jcl and the corpus already contained them; the rest is GITGALAXY-forged or unlicensed. data/jcl/SOURCES.md now documents this and what unblocks it (new licensed mainframe repos in gitgalaxy/data).

Tooling

.claude/skills/expand-language-coverage/ — a Claude Code skill packaging the
repeatable process these expansions used: read the target language's detection
standard, survey the pool for dialect-correct material (tools/survey_pool.py),
design a folder set and get it approved, stage each folder with provenance
drafting (tools/stage_folder.py), regenerate PROVENANCE.json, update every
SOURCES.md, open a PR. Hard rule: real source only, never synthesized; cap a
category and say so when the pool can't support the target size.

Provenance

data/PROVENANCE.json regenerated deterministically. exact-confidence repo-folders: 38 → 102; unknown 27 → 29
(embedded_python/meow_turtle, assembly/raspberrypi_baremetal). Full per-folder attribution (upstream URL, commit,
license, the specific parsing boundary each folder tests) is in the per-category
SOURCES.md files, indexed from the root SOURCES.md.

For GitGalaxy maintainers

Per RELEASING.md: regenerate the golden-master
fixtures and the tri-comparison / tree-sitter-accuracy artifacts against this tag
(docs/self_scan/BUMPING_THE_CRUCIBLE_PIN.md), then bump LANGUAGE_CRUCIBLE_REF

• tests/_crucible_pin.py's PINNED_TAG to v1.2.0.

Full diff: v1.1.0...v1.2.0