Surface the search filters as controls, extract snippets in the engine, and fix the test suite on modern Node

[franciscoxc]

Cardinal's engine is excellent and almost none of it is visible: nothing in the window suggests that type:image, content:"…" or in: exist. This PR adds controls for the common filters, plus a fix for the test suite, which is currently red on any recent Node.

I maintain a fork, but this branch carries none of it: no bundle identifier, signing config, icon, version bump, README or CHANGELOG change. tauri.conf.json and every icon file are untouched. It is only the parts that seemed useful to Cardinal itself. Happy to split it into smaller PRs if that is easier to review.

1. The test suite is red on Node 22+ (7c8eedb)

51 tests fail with Cannot read properties of undefined (reading 'clear'). Node 22+ defines its own global localStorage accessor that returns undefined unless --localstorage-file is passed, and Vitest's populateGlobal skips copying a jsdom key when one already exists on globalThis unless it is in an internal allowlist — navigator is on it, localStorage and sessionStorage are not. So jsdom's storage never reaches window. CI stays green only because it pins Node 20, which has no such global.

The test script now passes --no-experimental-webstorage, and CI moves to Node 24 since Node 20 rejects that flag outright (and went EOL in April 2026).

Same commit fixes AGENTS.md, which names nightly-2025-05-09 while rust-toolchain.toml pins nightly-2025-12-11. Following the doc and putting that older toolchain on PATH builds fine from a warm target/, then fails as soon as a proc-macro has to be rebuilt: rustc 1.88 emits dylibs that recent macOS dyld rejects with mis-aligned LINKEDIT string pool. That cost me an afternoon.

2. Content snippets in the engine, and type:email (30ee8b3)

To show why a content: search matched, the app needs the text around the match. Rather than a second scanner in the app, content_snippet sits next to node_content_matches in search-cache and shares its chunked read and Rabin-Karp finder, so a snippet is found wherever the filter found a match. Case folding goes to a scratch buffer instead of in place, because the snippet is cut from those same bytes.

The highlighted terms come from content_terms_of_query, which walks the parsed query instead of re-parsing the text in the UI. Two behaviours fall out of using the real parser: a negated !content: no longer highlights anything, and content:"Bearer " keeps its trailing space.

type:email covers .eml, .emlx, .emlxpart, .msg and .mbox, with mail/message as synonyms. Apple Mail's own formats are the ones nobody finds by name.

3. Filters as controls in the search bar (f3f67b5)

The rule these follow: a control edits the query text, it does not hold filter state of its own. So there is no second filtering path to keep in sync, the search bar stays the single source of truth, and the syntax stays visible — the user sees type:image appear and learns it.

• File-type dropdown — Image, Video, Audio, Document, Email, Archive, Code, App, Folder. "Document" writes an OR of doc/pdf/presentation/spreadsheet rather than widening type:doc, which would silently change that filter for everyone.
• A "Contains" field writing content:"…", beside the one for names.
• A context column with the matching text from inside each file, term highlighted.
• Reorderable columns, dragging their titles. The order is stored, and repaired on read so a column added later still appears.
• The folder icon opens a folder picker. It used to fold the field away, which is not what an icon shaped like a folder promises; folding moved to a chevron beside it. This adds tauri-plugin-dialog with only dialog:allow-open.
• The scope placeholder names its default: "Search in… (whole disk)".

Each control reads back only the token it authored. A query it cannot represent faithfully — negated, duplicated, or inside a boolean group — reads as "custom" and is left untouched rather than misreported.

One thing to push back on: renaming "folder scope" to "search in" across the 15 locales is opinionated. It is confined to the i18n JSON files and trivial to drop if you would rather keep the original term.

Checks

• npm test — 34 files, 292 tests, default pool and parallelism
• npm run typecheck, npx prettier --check .
• cargo test -p search-cache — 880 + the integration suites
• cargo check for the app, with Cargo.lock regenerated on the current version

New tests cover the snippet scanner across a chunk boundary, type:email and its aliases, the query-token round trip for both controls, and the column-order storage.

[franciscoxc]

CI was red on two checks. One was mine, one was not.

Tauri fmt — my content_terms line needed breaking. Fixed in d1748ee.

Rust test — test_globstar_dedup_trailing_expansion is flaky on master, independent of this PR. It intermittently returns the search root itself (a) alongside its descendants:

left:  ["a", "a/a", "a/a/file.txt", "a/file.txt"]
right: ["a/a", "a/a/file.txt", "a/file.txt"]

Running the test binary 400 times on each tree:

tree	failures in 400
master (4c50734), unmodified	5
this PR's branch	7

Same rate within noise, and this PR touches no globstar code — its search-cache changes are the content-snippet extractor, content_terms_of_query, and the type:email group. Happy to open a separate issue for it if that is useful.