ci(security): add a supply-chain gate, checksums and an SBOM

The project shipped signed Android builds and unsigned desktop binaries
with no vulnerability scanning of any kind. Nothing checked the ~500
crate Rust graph or the JS packages against an advisory feed, and nothing
checked that what we redistribute inside an MIT bundle permits it.

The first cargo-deny run found eight vulnerabilities and one
unsoundness -- bytes, four in rustls-webpki, time, two in quick-xml and
rand -- every one of them closed by a `cargo update` nobody had a reason
to run. That update is in this commit; 740 Rust tests and clippy
-D warnings pass on the new lockfile.

Two structural fixes matter as much as the gate itself:

  - deny.toml scopes the graph to the targets we actually ship. Without
    it the Apple targets pull in plist -> quick-xml and report two DoS
    advisories against a crate that is in no binary we release. Ignoring
    those by ID would silence them everywhere, including where they
    would matter; scoping makes them correctly absent.

  - libmpv is pinned by rev instead of branch = "master". A branch means
    the revision is whatever Cargo.lock happens to hold and any
    `cargo update` silently substitutes new upstream code -- in the one
    dependency that is not from crates.io and that links a C library
    into the player. The rev is the commit already locked, so this pins
    current behaviour rather than changing it.

Licence findings are recorded rather than waved through. libmpv and
libmpv-sys are LGPL-2.1, satisfied here by dynamic linking against the
system library; deny.toml carries the two obligations that follow (keep
the linkage dynamic, ship libmpv's licence text with any bundle carrying
the .so). MPL-2.0 crates are file-level copyleft and fine unmodified.

Releases now publish SHA256SUMS (verified in-job with `sha256sum -c`
before upload) and a CycloneDX SBOM for both halves, so "does this
release contain <vulnerable crate>?" has an answer that is not "rebuild
the tag and re-resolve it".

Workflows pin jellytau-builder:2026.08 instead of :latest. While every
job said :latest, rebuilding the image changed what every build compiled
against, including rebuilds of old release tags.

Also folded in, because both were the same class of problem:

  - publish-docs.yml downloaded mdBook from GitHub releases into
    /usr/local/bin at job time -- a toolchain install in CI, which
    CLAUDE.md explicitly forbids, and a hard dependency on GitHub's CDN
    at publish time. It is in the builder image now.

  - extract-traces.ts only ever read .ts/.svelte/.rs, so every
    requirement implemented by *configuration* was invisible to the
    matrix that measures it. DR-205, DR-206, DR-207 and DR-215 all carry
    TRACES comments nothing read, and each counted as uncovered while
    being covered. Coverage was really 90%, not 88%; MIN_THRESHOLD moves
    to 89 accordingly. CI workflows stay excluded and there is a test
    saying why: traceability-check.yml quotes "a TRACES: comment" beside
    deliberately-undefined example IDs, which the extractor would read
    as real traces and then fail its own dangling-ID check.

Supply-chain requirement is DR-216.

🔴 The builder image must be rebuilt and pushed
(scripts/build-builder-image.sh 2026.08) before this reaches master --
the workflows now name a tag and tools that do not exist in the registry
yet.

docs/requirements.md

@@ -405,6 +405,7 @@ Internal architecture, components, and application logic.
 | DR-213 | Containerised builds hand their artifacts back to the host user. The compose services bind-mount the repo and run as root — their caches live at `/root/.cargo` and `/root/.bun`, so a non-root container user cannot write them — which leaves root-owned files accumulating in the developer's working tree: 11,124 of them when this was found, enough that `cargo clean` and `scripts/clean.sh` failed with EACCES and a plain `cargo build` died part-way, since build scripts compile for the host and land in `target/debug` even during a cross-build. Ownership is restored at the end of each containerised build, reading the intended owner from the checkout so no uid needs plumbing through. Running the containers as the host uid is the tidier fix and remains open; it needs the cache volumes relocated off `/root` first | Tooling | - | Done |
 | DR-214 | The app identifies itself correctly everywhere a user or a package manager reads its name. `productName` was the scaffold's lowercase `jellytau`, which is what the Android release build showed under its icon and what the deb/rpm/NSIS bundles carried as their display name — invisible in development because `build.gradle.kts` overrides the label to "JellyTau Debug" for the debug build type, so the install a developer looks at daily was the only correctly-cased one. `mainBinaryName` pins the executable filename so nothing that resolves a path by name has to change. `strings.xml` moves into the canonical android tree, where `sync-android-sources.sh` already copies `res/values/*.xml`, so the fix survives regenerating `gen/`. Bundle metadata (publisher, copyright, category, descriptions, licence) was entirely absent, which is why the packages shipped with no maintainer or description — the hand-written Arch PKGBUILD and `.desktop` had all of it, so only the *generated* packaging was wrong | Packaging | - | Done |
 | DR-215 | Frontend test coverage is a ratcheted CI gate rather than a number nobody looks at. `test:coverage` had been configured since the suite was created and was silently broken: `@vitest/coverage-v8` resolved to 4.1.10, whose peer range pins `vitest` exactly, while `package.json` asked for `>=1.0.0 <5.0.0` and got 4.0.16 — so every invocation died on a missing `BaseCoverageProvider` export and no coverage figure had been produced in months. Fixing the range is half the requirement; the other half is that a measured figure that gates nothing decays the same way an unrun script does. Thresholds sit a few points under the measured result (statements 54.6, branches 48.7, functions 49.6, lines 55.1 when this landed) and only ever move up, matching `MIN_THRESHOLD` in the traceability gate and the eslint `--max-warnings` ratchet. The absolute numbers are held down by `.svelte` components, which this project deliberately does not test directly — the pattern is to extract the logic to a plain module and test that | Tooling | - | Done |
+| DR-216 | Dependencies are gated on known vulnerabilities and on licence compatibility, and the build graph is pinned to what is actually shipped. The project had no scanning of any kind: nothing checked the ~500-crate Rust graph or the JS packages against an advisory feed, and nothing checked that everything redistributed inside an MIT-licensed bundle permits it. The first run found eight vulnerabilities and one unsoundness — `bytes`, four in `rustls-webpki`, `time`, two in `quick-xml`, `rand` — every one closed by a `cargo update` nobody had reason to run. `cargo deny` (src-tauri/deny.toml) now runs in CI over advisories, licences, bans and sources. Two structural fixes matter as much as the gate: the graph is scoped to the targets actually shipped, so an advisory against an Apple-only path is correctly absent rather than ignored by ID; and the one git dependency (`libmpv`) is pinned by revision instead of by branch, since a branch means any `cargo update` silently substitutes new upstream code in the one dependency that is unsigned and links a C library into the player. Licence findings are recorded rather than waved through — `libmpv`/`libmpv-sys` are LGPL-2.1, which the app satisfies by dynamic linking, and that carries obligations (keep the linkage dynamic; ship libmpv's licence text with any bundle carrying the .so) | Tooling | - | Done |
 | DR-198 | The webview runs under a real Content-Security-Policy, and the asset protocol is scoped to the one directory it still serves. `csp` was `null`, which disables CSP entirely: any script that reached the web layer — through a future `{@html}`, a dependency, or a devtools paste — would have inherited the whole IPC surface, and with it the user's session. `script-src 'self'` (Tauri injects a nonce for SvelteKit's inline bootstrap script at build time, so no `'unsafe-inline'` is needed) plus `object-src`/`frame-src 'none'` and `base-uri 'self'` is the part that is genuinely restrictive. `img-src`/`media-src`/`connect-src` cannot be: the Jellyfin origin is typed in by the user at run time and is commonly plain `http` on a LAN, so they allow `http:`/`https:` — a wide grant for *data*, but one that still bars `file:`, `filesystem:` and scripting schemes, and leaves `script-src` untouched. `style-src` keeps `'unsafe-inline'` because Svelte compiles `style="…"` attributes (including `app.html`'s `display: contents` wrapper) into markup; this is safe only while no `<style>` element survives into `index.html`, since a nonce there would make Tauri's injection outrank — and therefore void — `'unsafe-inline'`. `worker-src blob:` and `media-src blob:` are hls.js: it demuxes in a worker built from a blob and attaches MSE through `URL.createObjectURL`. `asset:` and `http://asset.localhost` are the same protocol under the two naming schemes `convertFileSrc` emits (custom scheme on Linux/macOS, `http` host on Windows/Android); `ipc:`/`http://ipc.localhost` is the invoke transport, which would otherwise be blocked by `connect-src`. A run-time CSP naming the server origin exactly was rejected: Tauri computes the header from immutable config when it serves the HTML, so it would mean rebuilding config and reloading the webview on every server change, for a policy the user can already point anywhere. The asset-protocol scope narrows from `$APPDATA/**` to `$APPDATA/thumbnails/**` — since DR-137 moved downloaded media to the loopback server, `imageCache` is the only `convertFileSrc` caller left, so the database and the encrypted-token fallback file no longer sit inside the grant | Security | UR-012, UR-071 | Done |
 
 ---