git:20260725.1d299f1 to git:20260726.23bbca9

8 added, 5 removed. Audit A to A.

What this update changes in the checks

No check changed: every fixed check gives the same answer before and after this update.

A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.

# code_chunking/ — rules
- Tree-sitter chunkers, one per language (`python`, `go`, `cpp`, `csharp`, `rust`, `java`, `ruby`, `php`, `kotlin`, `js`).
+ Tree-sitter chunkers, one per language (`python`, `go`, `cpp`, `csharp`, `rust`,
+ `java`, `ruby`, `php`, `kotlin`, `swift`, `scala`, `lua`, `js`).
The authoritative reference for chunk depths and the full per-language
`chunk_type` tables is [`README.md`](./README.md) — read it before adding or
changing a chunker. This file is the short list of rules that bite.
## Conventions
- **Line numbers are 0-based here.** `CodeChunk.start_line`/`end_line` come
straight from tree-sitter and are **0-based**. The +1 conversion to the
1-based `CodeLocation` happens downstream in `dataset/gt_locate.py`
(`_chunk_to_code_block()`) — do **not** add an offset inside a chunker.
- **`.c` files use the `cpp` chunker.** There is no separate C chunker; the
language→chunker map sends `.c` to `cpp`. Forgetting this produces empty
`code_blocks` (the bug that silently broke redis/jqlang instances).
- - **C#, Java, Ruby, PHP, and Kotlin are chunker/GT/agent-only for now.** The `.cs`,
- `.java`, `.rb`, `.php`, and `.kt` chunkers are registered for tree-sitter chunking
- and ground-truth extraction, but graph/LSP/core backend support is still
- intentionally unset in the language registry.
+ - **Graph backends are per-language — check `codenib/languages.py`, don't
+ assume.** C#, Java, Ruby, PHP, Kotlin, and Scala have *active* SCIP cold-start
+ routes in the language registry (`scip-dotnet` / `scip-java` / `scip-ruby` /
+ `scip-php`), most with an LSP fallback (`graph_route='lsp'`); Ruby is also
+ core-decoder accelerated. Only Swift and Lua are chunker/GT/agent-only (no
+ graph/LSP backend registered), and C/C++ cold-starts via clangd, not SCIP.
- **Chunk depth** (`chunk_depth` on `BaseCodeChunker`): L0 = whole file
(skeleton when `skeleton_mode`), L1 = top-level symbols, L2 = nested
methods/members (default; with `l2_level_exclusive=True` the L1 containers are
dropped, keeping only their L2 members).
- **Symbol chunk types** (`SYMBOL_CHUNK_TYPES`): `function`, `method`, `class`,
`module`, `struct`, `type`, `interface`, `object`, `enum`, `trait`, `impl`, `var`,
`const`, `static`, `declaration`, `macro`, `variable`, `record`, `property`. Per-language L1
var/const kinds: Go `var`/`const`; Rust `const`/`static`/`type`; C++
`declaration`/`macro`; C#/PHP/Kotlin `property`; Java `record`; Ruby `module`;
JS/TS `variable`.
## When adding a language / chunk type
- Add it to the per-language table in `README.md` in the same PR.
- Heavily-templated C++ headers (e.g. fmtlib/fmt) confuse the chunker — use
`redis/redis` (plain `.c`) for C/C++ test coverage instead.