feat(player): seek strategy follows what the engine says it can do

DR-246. The strategy used to turn on `is_hls` and `use_html5`, decided in a
command handler on behalf of engines it does not own. That is how "who
renders" came to mean "how do I seek", and why a transcoded seek silently did
nothing the moment native video changed the renderer (DR-238).

Engines now declare `Capabilities::seeks_transcoded_in_place` — true for
hls.js, which seeks within the VOD playlist it was handed and lets the server
catch up; false for mpv, whose HLS demuxer cannot make the server transcode
from a new offset. The command asks whichever engine is rendering. Adding an
engine no longer means editing a shared truth table.

The item's transport is not read at the seek site any more; the compiler
flagged it unused, which is the URL-shape input finally disappearing.

A deviation from the spec, recorded deliberately: it called for the engine to
own the decision outright. It cannot. Re-negotiating a stream needs the
repository, which sits above the engine, so the engine states the ability and
the caller acts on it. That still removes the defect — nobody guesses on
another component's behalf — without pretending an engine can reach upward.

Also fixes a latent race in the conformance suite, found by running it: the
seek case asserted immediately, which passes on an engine that records the
target when it accepts a seek and races on one that waits for the decoder to
move. `Harness::await_seek` polls instead, the way the Android suite already
did. It failed with machine load rather than with the code, which is the kind
of test that teaches people to re-run until green.

  MpvPlayer     9/9
  LegacyPlayer  8/9 - still only the mute/rate gap in the old trait

789 tests, clippy -D warnings clean with and without the feature.

docs/requirements.md

@@ -441,6 +441,7 @@ Internal architecture, components, and application logic.
 | DR-243 | Every engine passes one conformance suite, and a `FakePlayer` implements the contract deterministically. The suite is written before the second engine so it cannot encode whatever the first happened to do, and it drives readiness through a harness rather than sleeping. `FakePlayer` models the one behaviour that matters — opening is not instantaneous — so the load/seek race can be expressed on purpose, and lets the controller, queue, autoplay and session logic be tested with no engine at all | Player | UR-081 | In Progress |
 | DR-244 | `MpvPlayer` implements `MediaPlayer` over libmpv, applying the start position at load time via mpv's own `start` option rather than seeking after an asynchronous `loadfile`, and holding a seek that arrives during `Opening` until the file loads. A standalone `player-conformance` binary runs the suite against it with audio and video routed to null, so a wrapper is verifiable without building or launching the app | Player | UR-081, UR-040 | Done |
 | DR-245 | `PlayerController` holds a `MediaPlayer` rather than a `PlayerBackend`, and every engine reaches it through that one contract — `LegacyPlayer` carries the not-yet-ported ones across unchanged, so the port swaps a seam rather than four implementations. Loading an item is now a single `open` carrying its start position, and the controller maps the engine's `Phase` back onto `PlayerState` using the queue, so nothing outside changes. `LegacyPlayer` drives the old `PlayerBackend` through the `MediaPlayer` contract, so engines not yet ported keep working during the migration and the two designs can be compared on one engine and one file. It reproduces the old load-then-play-then-seek sequence faithfully rather than a fixed-up version, because making it pass would defeat its purpose | Player | UR-081 | Done |
+| DR-246 | The seek strategy turns on an ability the engine declares, not on the container the stream arrives in. `Capabilities::seeks_transcoded_in_place` is stated by each engine — true for hls.js, which seeks within the VOD playlist it was handed; false for mpv, which cannot make the server transcode from a new offset — and the command asks the engine currently rendering instead of inferring from `is_hls` and `use_html5`. The item's transport is no longer read at the seek site at all. Re-negotiating a stream needs the repository, which sits above the engine, so the engine states the capability and the caller acts on it rather than the engine owning the whole decision | Player | UR-040, UR-081 | Done |
 | DR-247 | ExoPlayer can be told where to start. `JellyTauPlayer.load(url, mediaId)` had no way to express a start position, so every caller loaded and then seeked; the position is now handed to ExoPlayer with the media item via `setMediaItem(item, startPositionMs)`, and the two-argument form delegates to it. Running the conformance cases on a device also settled which half of DR-241 was engine-specific: ExoPlayer already queues a seek issued before `prepare()` completes, so it never had the lost-seek defect mpv did — only the missing vocabulary for a start position | Player | UR-081, UR-005 | 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 |