[Dr-Amp]

Part of #918 ("Opus direct play up to stereo").

LG's buffer feed decodes Opus (OPUS + contents.opusInfo, as moonlight-tv's ss4s feeds it), but DP_AUDIO_CODECS never offered it, so every Opus track went to the server and Force Direct Play refused it.

Change

• devcaps: opus is a pipeline-only audio codec: admitted only on a device table that was actually read, never on the assumed fallback, in Auto and in Force Direct Play (forced_feeds_audio) alike.
• Auto takes stereo only (OPUS_AUTO_MAX_CHANNELS = 2), whatever a table row claims, following the C1 measurements in FLAC: decode to PCM in the app, up to stereo #917 (5.1 stalls once packets pass ~1500 bytes, with nothing in the log). The cap rides audio_channels, so the profile's audio.channels limitation hands surround Opus to the server and the planner refuses it locally. Force Direct Play keeps no ceiling, as for every codec.
• engine: an OPUS Load carries opusInfo {channels, sampleRate: 48, streamHeader: base64(OpusHead)}: mapping family 0 for 1–2 channels, Vorbis family 1 for 3–8 (libopus's surround layout; the 5.1 header matches a real track's CodecPrivate byte for byte). A width it cannot describe refuses the Load instead of guessing.
• transcoder: opus joins the copy-only target lanes, last, so it is never the encode head.
• ci/list-caps.ini classes the new constant (not-a-list).

Verification

Host: CachyOS x86_64, pinned toolchain 1.99.1, on 3b7c6e28 (main e843b99f):

• check-cargo-lint: pass. check-nodef: pass. check-python-harness: Ran 542 tests … OK (skipped=1).
• check-cargo-unit-default / check-cargo-unit-hostsim: every crate passes except plx_media's two curlio::keymode_tests (1709 passed; 2 failed / 1870 passed; 2 failed). The same two fail on unmodified main e843b99f on this host (6 passed; 2 failed, filtered run): libcurl 8.22 here, not this change.
• check-python-rest: one failure, ci/test_disk_preflight.py ("only 2.0 GiB free for ~/main/rust-modules/target"), identical on unmodified main: the host's disk.
• Not run here: the ARM cross-build (CI's job) and check-replay (no Mac; no screen shape changes).

Device: LG C5, webOS 10.3: an AV1 + Opus 5.1 MKV direct-played through the family-1 header, with seeks and next-episode autoplay. That run predates the stereo cap, so it was Auto; it is now Force Direct Play's path. The stereo (family 0) header is covered by unit tests but was not played on the panel.

Known gap: a native switch between two Opus tracks of different widths keeps the first track's channel count (route::stream_achannels).

Signed-off-by: Dr-Amp dr_amp@users.noreply.github.com

[GLinnik21]

Thanks @Dr-Amp, Opus direct play is a nice one to tick off #918, and the byte-for-byte check of the 5.1 header against a real CodecPrivate is great. Merged.

Small cleanups I pushed before merging, behaviour the same:

• with_opus_info uses the shared plx_base::b64::encode instead of a local base64 helper. Your RFC 4648 vector test moved next to it in base/src/b64.rs.
• in devcaps parse(), the Opus Auto ceiling is one entry(..).or_default() + min_nz, and the table_gated closure became direct PIPELINE_ONLY_AUDIO.contains checks.
• main merged in clean.

tiper's #1034 builds on this, so it goes in next.