docs(plans): promote the language registry, cost approved

Moves editor-language-registry from ideas/ to plans/ and rewrites the header
from "is this worth it" into a spec. The measurements stay, as the accepted
cost rather than the open question: app start unchanged, +40K on the editor
chunk, +824K in the .app across 94 lazy chunks.

Says what "automatic" means, since that was the question behind the approval:
matching is by FILENAME, which is extensions AND whole-filename patterns
(Dockerfile, Makefile, justfile, .env.production), so `languageIdForPath`'s
special-case list goes away with the catalog. A manual Set-syntax pick still
wins, and the content sniffer still answers when the filename says nothing —
which is every scratchpad, by definition.

Also pins the two things that keep this honest: our Makefile and proto3
grammars are registered ALONGSIDE the registry rather than replaced by it
(nothing upstream has a Makefile grammar), and `lib/languages.ts` may not
import language-data, since the palette imports it into the main chunk — a
source-level test, like cspGuard.

No `planned` issue opened: the issue rule says never to open one for a doc I
just wrote. Ready to open it on your word, and it needs a README Planned
bullet, which is maintainer-only.

Refs #244

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F2JzhUsFa9YLjBBvUJQGYj

Author: simion

docs/ideas/editor-language-registry.md

@@ -0,0 +1,121 @@
+# Language support from the filename, with no per-language edits
+
+**Approved.** Cost accepted (measured below). This replaces the hand-maintained
+language catalog with CodeMirror's published registry, so opening a file of
+*any* language it knows lights up without anyone editing termic first.
+
+Today adding a language is two edits — an entry in `lib/languages.ts` and a
+`case` in `lib/languageExts.ts`. That is how Swift and Groovy/Gradle went in.
+After this, neither file has a list to add to.
+
+## What "automatic" means, precisely
+
+Matching is **by filename**, which is both halves of what that implies:
+
+- **Extension** — `.swift`, `.gradle`, `.php`, `.zig`.
+- **Whole-filename patterns** — `Dockerfile`, `Makefile`, `justfile`,
+  `.env.production`. `LanguageDescription.matchFilename()` handles both, so the
+  special-case list in `languageIdForPath` goes away with the catalog.
+
+Two things stay in front of it, unchanged, and the precedence order in
+`effectiveLanguageId` does not move:
+
+1. **A manual "Set syntax" pick** still wins over everything.
+2. **The content sniffer** (`lib/detectSyntax.ts`) still answers when the
+   filename says nothing — an extension-less file, and every scratchpad
+   (GH #244), which by definition has no filename at all.
+
+## The registry
+
+`@codemirror/language-data`: ~150 `LanguageDescription`s, each carrying its
+extensions, filename patterns, aliases, and a lazy loader.
+
+```ts
+const desc = LanguageDescription.matchFilename(languages, "build.gradle");
+const support = desc && await desc.load();   // dynamic import, code-split
+```
+
+**Our two extras are registered alongside it**, not replaced by it: nothing
+upstream has a Makefile grammar (`lib/makeMode.ts`), and the shipped protobuf
+mode predates proto3 (`lib/protoMode.ts`). The end state is "the registry plus
+our two", expressed as `LanguageDescription.of(...)` entries prepended to the
+list so ours win on a tie.
+
+## Accepted cost
+
+Measured by wiring the registry into `EditorPane` behind a live call site and
+diffing `npm run build` against the same build without it. (A tree-shaken probe
+measures nothing — the first attempt reported an unchanged bundle because
+rolldown dropped the reference. Re-measure that way if these numbers are ever
+rechecked.)
+
+| | before | after |
+| --- | --- | --- |
+| main `index` chunk | 2300K | **2300K** (unchanged) |
+| editor chunk (first file opened) | 604K | **644K** (+40K) |
+| total `dist/` | 7208K | **8032K** (+824K, 94 lazy chunks) |
+| new npm deps | — | 9 (`language-data` + 8 Lezer grammars we lack) |
+
+App start pays nothing, the first opened file pays 40K for the index, and the
+`.app` carries ~0.8MB of grammars that are only ever *loaded* for a language
+the user actually opens.
+
+**Hold app start at zero.** The registry may only be reachable from the lazily
+loaded editor/diff panes. `lib/languages.ts` is imported by the command palette
+and the breadcrumb, i.e. the main chunk, so it must not import
+`@codemirror/language-data` — pin that with a source-level test, the way
+`cspGuard.test.ts` pins the CSP.
+
+## The work
+
+**1. Async loading.** The awkward part. `langForId` returns an `Extension`
+synchronously today and the editor mounts with it; the registry returns a
+promise. So the editor mounts with no grammar and reconfigures its language
+compartment when the load resolves. The compartment is already the live-switch
+mechanism (Set-syntax uses it), so the machinery exists.
+
+It needs a **cancellation guard**: bump a generation counter per mount/path
+change, capture it before the await, and drop the result if it no longer
+matches. A fast tab switch must never land a resolved grammar in a view that
+has since been rebuilt — that is the async-mount race the no-StrictMode rule
+exists for, and it gets its own test.
+
+**2. Ids become registry names.** `EditTab.syntax` is session-only today, so
+there is nothing to migrate — but fix the contract now, because the scratchpad
+plan persists a pick in its index. Use the registry's `name` ("TypeScript",
+"Makefile"); `languageLabel()` keeps rendering an unknown id verbatim so a
+stored name that later disappears degrades to plain text instead of blank.
+`lib/detectSyntax.ts` returns those names too.
+
+**3. `lib/languages.ts` gets thin.** It keeps `PLAIN_TEXT`,
+`effectiveLanguageId`'s precedence, `languageLabel`, and the note pointing the
+LSP work at it as the one place a `languageId` is decided. The `LANGUAGES`
+array and `languageIdForPath`'s rules are deleted, not ported.
+
+**4. The picker sources from the registry.** `SyntaxPalette` maps
+`LanguageDescription[]` to rows (name + `alias` for the fuzzy search), keeping
+Plain Text pinned first. That is a straight simplification.
+
+**5. `lib/languageExts.ts` loses its switch**, keeping only the two custom
+descriptions and the registry lookup.
+
+## Testing
+
+- **Unit:** the filename cases the current suite already pins must survive the
+  swap — `Makefile`, `Dockerfile.dev`, `justfile`, `.env.production`,
+  `build.gradle` (Groovy) vs `build.gradle.kts` (Kotlin), `README.MD`. Plus the
+  new invariant: nothing in the main chunk's import graph pulls in
+  `language-data`.
+- **Async race:** mount, switch path before the load resolves, assert the stale
+  grammar never lands.
+- **e2e:** the existing `editor.e2e.ts` syntax cases stay as they are — they
+  assert *rendered tokens*, so they prove the async path really reaches
+  CodeMirror. Add one case for a language nobody hand-added (PHP or Lua): open
+  the file, get highlighting, with zero termic-side registration. That is the
+  whole feature in one assertion.
+
+## Out of scope
+
+Highlighting inside terminal output; anything LSP (`docs/ideas/lsp.md` owns
+that, and should read its `languageId` from the same place); shipping grammars
+the registry does not include.