Consume langbank instead of carrying src/profiles/

[zmaril]

The four-repo decomposition named this boundary at design time: langbank owns "what is this language", entl owns "what is in this codebase." src/profiles/ was the part of entl on the wrong side of it — the module langbank was lifted from — and every day both existed was a day their answers could drift apart.

86 files changed, 189 insertions(+), 3,594 deletions(-)

Nothing is re-exported

Code that needs a LanguageProfile, an EcosystemProfile or a detection takes it from langbank directly — entl-codebase, entl-github and entl-tree-sitter each carry the dependency and name the types where they use them. entl's public surface is what entl owns: walking, inventory, manifests, packages, workspaces. langbank types appear in it only where a boundary type genuinely crosses (FileEntry.language is a langbank::LanguageDetection; an Artifact names a langbank::ArtifactId). A consumer that wants langbank's vocabulary depends on langbank, not on entl's copy of its names.

PackageId/WorkspaceId — genuinely entl's — are minted by entl's own private string_id!, as before this change. langbank exports its macro for consumers, but two newtypes don't justify a structural dependency; the convention matches by construction either way. Recorded inventories read back exactly as written (transparent serde on both sides).

Behaviour change, measured rather than assumed

Walked three repositories before and after, diffing every file's detected language:

walk	files	gained	lost	changed
entl itself	135	18	0	1
langbank	2,140	3	0	0
treebank	406	140	0	0

Gains are what absorption bought: Cargo.lock is toml, .scm queries are scheme, treebank's 112 patch files are diff. The one changed answer is tsconfig.json: json → json-with-comments — correct (it's JSONC; linguist maps the filename), so the old answer was the bug. entl grows 8 → 827 languages, 5 → 36 ecosystems.

Carried across the rebase

main gained LanguageRole::expects_parser_pack() while this branch was open — on the very file it deletes. It survives as an extension trait in entl-tree-sitter, original reasoning verbatim. That's its right home rather than a workaround: whether a parser pack exists for a role is a fact about that crate's pack format, not langbank's judgement about languages.

Removed with the module

verbosity (1,334 lines) — langbank dropped it as unreproducible (generator and method write-up both already lost). Nothing outside this crate consumed it; its seven tests go too. langbank's docs/verbosity.md records what it was and the git show that reads it back.

Tests

Two tests pinned the old world exactly — five ecosystems, three conventions — and now assert the floor rather than the ceiling, so an absorption in langbank doesn't turn into a red build here. profile_registration.rs still proves a downstream crate can register its own profiles, importing the registration types from langbank as a real downstream would.

181 passing, clippy -D warnings, fmt clean. langbank is public, so the git dependency fetches anonymously — no CI secret needed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HJZVnN8ShZktJ8VSSL3AjD

[zmaril]

Rebased onto current main, which had gained the anchoring/propbank design PRs since I branched — including one thing that was nearly lost:

LanguageRole::expects_parser_pack() was added to profiles/language.rs on main, and my deletion of that file would have dropped it. It survives as an extension trait in entl-tree-sitter, with the original reasoning carried verbatim — and that's arguably its right home rather than a compromise: whether a parser pack exists for a role is a fact about that crate's pack format, not langbank's judgement about languages. langbank's own rule is that data states facts and consumers apply policy; this is policy.

Post-rebase: 181 tests passing (main's new anchoring tests included), fmt/clippy clean, and the treebank walk diff re-run byte-identical.

The LANGBANK_READ_TOKEN secret note stands — CI here can't fetch the github.com private repo without it.

[zmaril]

Correction to the PR body's warning: no secret is needed. langbank is public — I wrote the CI auth plumbing on the assumption it was still private, then noticed the first CI run passed without any token. The LANGBANK_READ_TOKEN step is removed; the fetch is anonymous and worked in CI already. git-fetch-with-cli stays, as ordinary hygiene rather than a credential workaround.