Repository navigation
Merge upstream 2026-11-14 - #36
Merged
Merged
Conversation
…he release that uploads it (plxnative#1370) A crash, panic, freeze or exit() record is uploaded at the next launch, which after an update is a newer build, and every fallback report was stamped with the uploading build's release. Sentry PLX-NATIVE-26 was a 0.9.0 shutdown crash (its dist is 0.9.0's build id) filed as a 0.10.0 crash. - stamp_release, shared by all four bodies: when the record's img: build id differs from the running build's, the report carries no release; dist stays the crashed build's id, and the tags crash.uploaded_by (the sending app version) and crash.build_id (the crashed build, since Sentry ignores dist without a release and a panic body has no debug_meta) say what happened. Same build, no image line or an unreadable running id: stamped as before. - Io::queue_report handed the bodies the record's own id as the running id; body construction is now report_body(), shared by the queue and the tests, and receives current_build_id. - The consent preview's panic sample shows the straddled shape; PRIVACY.md says when release is left out and names the two tags; the set_release doc names the exception. No record carries the crashed build's version (the image marker is written in src/main.c before the release is set), so these events are filed under no release rather than under the right one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lxnative#1331) `EXPORTED_STRATA` (`play-uri/src/backend.rs`) was empty, so the Experimental Playback engine fell back to Standard on every webOS 5+ set. It now lists firmware major 5. On a webOS 5.x television, a viewer who chose Experimental now plays an original the route accepts (one audio track, no subtitle stream, quality Original or Auto on a local link, a file Standard would direct-play) through the TV's own media pipeline in the exported window, with this app's player screen on top. Such a set's boot line reads `uri: product binding=Exported { canvas: (1920, 1080) } mode=2 ... major=5 evidence=table`. Everything else plays as before, including every other webOS 5+ major and the 4.x ACB path. The evidence meets all three parts of the admission rule in `docs/uri-mode.md`, from an LG OLED CX on 5.6.2 (plxnative#1064): - (a) the 5.5.0 and 5.6.2 images read `option.windowId` at load; - (b) a product run of a single-audio original on a debug build with the override armed: placed, correctly letterboxed picture, sound, pause, resume, seek and Back; - (c) a drive probe on the plxnative#1314 build that printed `pause=froze pause_basis=paused_event_and_resume`, `seek=hit` and `audio_at_load=0`. Rule (c) gains two amendments, named in the row that uses them: - (1) a run made before the current grader may stand on that grader's verdict on its recorded numbers, replayed in a unit test (unused by 5.6.2); - (2) a control that failed is acceptable only when the route now sends that case to Standard. This covers the first run's audio-track switch: the service orders tracks differently from the container, and since plxnative#1314 a multi-audio original starts in Standard, as the set has since logged. The composition test pins major 5 to `Exported` and every other major to `Unbound`, and the boot-line test pins `major=5 evidence=table`. The 5.6.2 Results row quotes the tester's lines; the 10.x drive probe is recorded without an entry. Still untested: subtitles and HLS on the URI path, other 5.x minors, and a build with this table entry on a 5.x set. Part of the webOS 5+ URI work tracked in plxnative#1064; does not close it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…celled, not as a direct-play refusal (plxnative#1372) When the detail page halts a trailer preview while its resolve is in flight, the landing that arrives afterwards starts no engine. Since plxnative#1274 that landing finished its request metric as Failed(PreviewNotDirect), so the app's attempt metrics recorded a direct-play refusal (and a HandledError notice) for what was a cancellation. It now finishes Cancelled, the same finish_metric_request + sync_request_metrics pair the viewer's exit and shutdown use. Nothing outside the app saw the wrong outcome: the metric notices are a bounded local window, and the preview's Sentry report comes from report::tick_preview, which only fires on a genuine PreviewNeedsDirectPlay refusal (unchanged). The new hostsim test drives a halted landing through update() and asserts Cancelled with no HandledError. The PreviewReset drop guard moves to bridge_test_support so this test and deeplink_route_tests share it, and the process-wide preview machine is reset even when a test panics. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… Out is on PCM (plxnative#1375) A soundbar or receiver on HDMI (e)ARC can play Dolby Atmos / E-AC3 titles slow, low-pitched and drifting out of sync on some starts, while the TV speakers play them fine. The reporter of plxnative#1351 traced it to the TV's Digital Sound Out being set to PCM; Auto fixed it. `docs/troubleshooting.md` gains a section for that symptom: set Settings > Sound > Advanced Settings > Digital Sound Out to Auto or Pass Through instead of PCM. With PCM the TV decodes the Atmos track itself and sends PCM over (e)ARC, which is LG's path and not something the app can change. `docs/playback-settings.md`'s Digital Sound Output sentence links to it. Docs only. Closes plxnative#1351 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lxnative#1373) A trailer preview only plays a plain direct play, and its refusal reached Sentry as preview_not_direct (0.9.0: unspecified, PLX-NATIVE-1G, 192 events / 49 users) without saying which condition refused it. - PlayVerdict::PreviewNeedsDirectPlay { reason }, chosen by the pure player::preview::refusal as the first failing condition: no_part, not_direct (no direct play offered, relay and quality policy included), adaptive (a direct play was offered and Auto chose HLS). plan.rs snapshots the direct-play offer before Auto's override, so Auto over a server transcode still reads not_direct. - Sentry tag playback.preview_refusal, only for kind preview_not_direct; fingerprint unchanged. - Consent legend: a row for the three reasons (12 locales), and the failure-kind row is built from FailureClass::ALL; it had omitted six kinds that reach reports (preview_not_direct, jail_missing_rtkmem, load_timeout, release_timeout, resolve_failed, server_stopped). - PRIVACY.md names the tag and its values. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e stops splitting by instruction (plxnative#1374) MainThreadFreeze reports (PLX-NATIVE-24/-25) arrived with no app frame and nothing saying what the main loop was doing, and a no-app-frame freeze was fingerprinted on a firmware-library offset, so one stall split into several issues. - A process-wide AtomicU8 holds the loop's current leg in every build: LoopPhases::enter stores the seven heartbeat legs; diag::spans::span stores the four span legs (scen, pump, land, dframe) and restores the enclosing one on return and unwind. Relaxed stores; threadcheck labels unchanged. - The watchdog's observer thread appends freeze-at: seq=<n> leg=<word> to the crash log when it sends the freeze probe, so an unrecovered freeze still names its leg; the event-log line names it too. - The importer matches freeze-at: by seq within the same process, in any order against the handler's lines, accepts only a word from the fixed table, and tags freeze.leg. - Fingerprint ["freeze", site, leg] (leg only when present); with a leg and no in-app frame the site is the pc/lr module basenames without offsets. Records without a leg keep their old fingerprint. - PRIVACY.md names freeze.leg and its eleven words. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…readable (plxnative#1376) When `/var/run/nyx/os_info.json` could not be read at boot, the app treated the firmware as unknown: every rule keyed on the webOS major ran as for firmware 0, Diagnostics said "webOS unknown", and the release sent to the Plex server fell back to the literal "4.5". A 6.5.3 set in plxnative#1351 hit this on one boot. `/usr/lib/os-release` is readable there (the developer-mode jail mounts `/usr/lib` read-only), so `webos.rs` now falls back to it when the nyx read fails. A pure `os_release_info` takes the release from `VERSION_ID` and the major from its first component. It requires `ID=starfish`, so a Linux simulator host's own os-release is never read as a firmware. It also refuses a major below 5: every unpacked image from 5.5.0 to 11.2.0 ships the file and no 3.x or 4.x image does, and a major 1 to 4 would flip the credential default, the VP9 10-bit rule, the DV platform gate and the old-ABI seek. A pure `firmware_from` consults the file only when the nyx read failed. Codename, API, name and build stay empty rather than invented, so the crash fingerprint never splits one set into two builds. The nyx verdict is unchanged: it still reaches the event log and the reports' `hardware.probe` field, and a fallback boot adds `webos: release=<r> major=<n> (from /usr/lib/os-release; nyx os_info.json unreadable)`. Diagnostics' head line drops an empty `api` label. Comments that equated major 0 with "os_info unreadable" now name both sources. Closes plxnative#1351 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…what a Flutter for webOS package is
The Seller Lounge user guide states the native rule in print ("If you do not
have a separate standing contract with LGE, select web > All"), and lists
per-chipset Alpha Test and Cloud Test Lab handling for native files.
A Flutter for webOS app built with LG's toolchain is an ordinary native ipk
(the portal reads File Type: Native); there is no flutter package type
(ares-package 3.2.6 rejects it); and its reach is webOS 26+ because LG's
prebuilt engine and the NDK 11.2.0 sysroot require glibc 2.38, not because
of the CPU target, which matches ours.
Also records LG's ACG table (required from webOS 27, requiredACG in every
manifest) as the next seam for pkg/appinfo.json, the NDK licence's scope in
the open question on LG's terms, and two new open questions: whether this
binary runs on webOS 26, and what LG's native-partner contract involves.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P6ZnFcPGgZcGvchojYjFn1
Takes the ed25519-compact 2.6.1 bump from the rust-minor-patch group and keeps zip at 9.2.2: zip 9.5.0 makes constant_time_eq and zeroize non-optional, which pushes the graph from 93 to 97 packages (over the build budget) and needs new licence rows. Moves the fuzz crate's ed25519-compact pin to match plx_data and re-pins the notices and the read-only proof certificates to the new lockfile. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01P6ZnFdEVcdu4E8TSFndps7
Pins rust-toolchain 1.100.0. The 1.100.0 std lock gains find-msvc-tools (a build-time dependency of cc, not part of the ARM binary), listed under std_not_shipped; the toolchain snapshot and THIRD-PARTY-NOTICES.md are re-derived, and the notices tests' next-pin fixture moves to 1.101.0. make build-bench, same machine back to back, 1.99.1 vs 1.100.0: every scenario equal or 0-6% faster, same rebuild fan-out, target dir 2438 MB -> 2423 MB. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01P6ZnFdEVcdu4E8TSFndps7
## Summary - Finish the UI ownership migration: remove legacy navigation/press publications, the shadow app press, and the global idle/damage adapters. - Route keyboard/pointer commits, deferred rendering wakes, publication damage, and actual-swap acknowledgement through the owning dispatcher. - Prevent premature settled screenshots, preserve late draw work, and advance the developer replay census to AppFrameV5. - Retain migration regressions and add independent-owner, source-pass, nested Settings and adapter-resurrection checks. ## Validation Passed locally through WSL on the pinned Rust toolchain: - `make check-cargo-unit-default` - `make check-cargo-unit-hostsim` - `make check-cargo-lint` (including feature checks) - `make check-nodef` - `ci/check-deps.sh`, `ci/check-statics.sh`, and new negative gate self-tests This PR is also intended to run the complete CI equivalent of `make check` and the ARM cross-build in clean runners. Local full Python validation was limited by Python 3.10/Windows checkout constraints; 28 SessionMachine shape findings were reproduced unchanged on the HEAD baseline. The local ARM compiler is not installed. No native-device or pixel-parity result is claimed. ## Compatibility Old developer recordings are explicitly rejected by the AppFrameV5 state fingerprint rather than replayed under different press semantics.
…t controls (plxnative#1380) a_peer_that_never_pauses_cannot_hold_off_a_cancel asserted that run() returned within 1 s of the cancel. That measured its own callback: the pass in flight still delivers up to PASS_BUDGET (about 1045 of the test's frames), each callback sleeps 200 us, and on the build bench's 3-core virtual macOS runner a 200 us sleep takes over 1 ms, so the test failed every day from 2026-11-10 and the bench's unit-suite row recorded nothing. The test now bounds what the transport controls: - messages delivered after the cancel are at most one pass, (PASS_BUDGET + READ_CHUNK) / frame + 2; a pass with no budget delivers thousands; - the time outside the callback between the cancel and the return, the going-away close included, stays under 1 s. It passes with a 200 us callback (0.33 s) and a 1.5 ms one (2.28 s), which fails the old assertion. With PASS_BUDGET = 1 << 40 it fails on the message count. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…leep (plxnative#1382) the_window_reread_waits_for_the_outgoing_stop_within_its_bound started its clock inside the spawned reader thread, while the 250 ms sleep before stop.finish() runs on the test thread right after the spawn. A reader scheduled late began timing after the sleep had started, so a read that correctly waited for the stop measured under 250 ms. CI's hostsim job failed on it on 2026-11-12 and 2026-11-14 ("it waited for the stop: 249.3ms"). The clock now starts before std::thread::scope, so a correct wait always measures at least 250 ms. The bound, the assertion and the window_reads() checks are unchanged. A 50 ms delay injected at the start of the reader fails the old code (204 ms) and passes the new one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… first, locale-proof sched-trace flavour check (plxnative#1384) The host harness job grew from about 240 s to 375-470 s and hit its 10-minute limit six times on 2026-11-12/13. 2647 of the module's 2678 summed test-seconds are DepGates, whose ~75 gate runs each execute all of ci/check-deps.sh; the 11-13 jump was the wall rule (ci/check-ports.py, plxnative#1358) lexing ~690 files in pure Python in every run. - ci/check-ports.py lexes only files whose source names Instant, SystemTime or elapsed, and computes a scope's wholly-test set only when that scope has a wall site (4.1 s -> 1.6 s CPU per run). Output is identical for every file that contains a wall word. An unlexable file with a wall word still fails unless it is wholly-test, as before; one without a wall word is left to ci/check-list-caps.py, which lexes every production file and fails closed. ci/test_check_ports.py pins both. - tests/test_harness.py: the two ui-owner loop tests plant every variant in one copy and require each by file name in one run, and the two tests that grade the checkout share one run behind a lock (75 -> 70 gate runs per pass). Nothing is skipped. - ci/check-python-steps.txt starts the nine slowest steps first: ci/test_check_machine_shape.py, one 225 s process, used to start 190 s in and end the job at 413 s; a list schedule of the measured durations gives 317 s for the new order. - tools/tv-sched-trace.sh spells its lower-case class out: under bash 3.2 with en_US.UTF-8, `[!a-z]` matched "Debug" as lower case, so --app=Debug exited 0 instead of 2 and failed the build bench's check row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…the CPU (plxnative#1385) The slowest unit tests were CPU-bound debug-build work, not waits. Each now does the same checks for a fraction of the CPU: - telemetry byte cap: the crash body size was found by framing a ~100 KB record about 2700 times; it is now derived from the empty record's frame overhead, with two asserts pinning it to encode() (28.4 s -> 0.08 s). - machine Canon: the state-hash encoder feeds bytes through an index loop instead of a per-byte iterator call; the bytes hashed are identical, pinned by a test against a byte-at-a-time reference and a constant from the old encoder. The screens focus walks spent 94% of their time here (the related-row walk 23.5 s -> 11.6 s). - ui Dispatcher::frame_with: the debug hash pair around the prepare step runs only when the prepare pass runs; with draw=false the only call between them is rig.opaque_route, so it compared a state with itself. - ui tab-capsule contrast: the pill's ink regions are decided once per call, and a pixel's contrast is reused when its backing and ink equal the previous one's (12.4 s -> 3.4 s, 5.8 s -> 1.7 s). - appkit HUD title: the unpaused title is drawn once per language, slot and kicker instead of for every focused control and pop frame; the control-rect check still runs for each (12.7 s -> 0.25 s). - media ASS deep read: MAX_EVENTS + 300 events leave the same full window as 30 000 did before the rewind (9.2 s -> 1.3 s). A production mutation per crate still fails the changed test. machine.rs is pinned by the generated-enum certificate; the edit is outside the certified newtype! macro and its ten invocations (lines 1-881 unchanged), so ci/readonly_generated.py and ci/proofs/platform-generated-enums.json accept the new file hash. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI's host lint and unit jobs built every crate with full DWARF, which nothing in those jobs reads. The host-rust action now appends [profile.dev] debug = "line-tables-only" and [profile.dev.package."*"] debug = false to the runner's ~/.cargo/config.toml. Developer checkouts, the daily build bench and the ARM profiles are untouched (they do not use this action or do not inherit dev). rust-cache keys on the workspace's .cargo/config.toml, not the runner's, so without a new key the jobs restored main's full-DWARF cache as an exact hit, rebuilt every dependency and saved nothing, on every run. The cache step now carries prefix-key v1-rust-linetables, to be bumped whenever the appended block changes. Measured on a warm run (43558025990, attempt 2): the unit-default job took 360 s against 434 s on a warm main run and a 468 s median; its workspace units summed 389 s against 496 s (-22%); the saved caches are about 30% smaller. Hostsim and lint moved within runner noise. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stsim (plxnative#1381) The hostsim unit pass re-ran every crate of the default pass. It now runs HOSTSIM_SUITE, the 11 crates that declare a hostsim feature (plxnative-modules, plx_base, plx_machine, plx_platform, plx_gfx, plx_plex, plx_telemetry, plx_data, plx_session, plx_media, plx_play_feed). plx_net, plx_ui, plx_appkit, plx_screens, plx_media_codecs, plx_play_uri and plx_playback_contract still run in the default pass. Their own test sets are identical in both passes (same names and counts in CI run 43528579834). The old pass linked them against hostsim-built plx_base, plx_gfx, plx_platform, plx_media and plx_play_feed (steerable paths, larger glyph caches, the FFI host stand-in); none of their tests sets the environment or reaches the code that observes those differences. ci/test_test_crate.py pins that HOSTSIM_SUITE is an ordered subset of the unit suite, that the recipe names exactly it, and that every crate left out declares no hostsim feature and has no feature = "hostsim" cfg, so a crate that gains one must be added back. The CI hostsim job ran 238-332 s across the PR's heads against a 395 s median; the local build bench shows no regression. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…xnative#1386) Each of the ~70 DepGates self-tests ran all of ci/check-deps.sh (20-25 s of CPU) to prove one rule catches one planted violation; that was nearly all of the host harness job, which reached 375-720 s and timed out six times. ci/check-deps.sh takes --only a,b to run just those rules and --list-rules to print the 43 names. Each rule's body sits in `if want <name>; then ... fi # <name>` (or `want <name> && gate ...`), unindented, so a selected rule runs exactly the code of the full run; the three shared preludes are computed only when a selected rule reads them. With no argument every rule runs and the output is byte-identical to before. An unknown name, an empty list, a repeated --only or an unknown flag exits 2 before any rule runs, and every run checks that RULES and the guards agree in both directions. Every DepGates self-test declares its rules (@_grades) and runs only those; the verdict helper refuses a run in which a selected rule printed no verdict line, and a test may only assert on rules it selected. Two whole-script tests still run every rule, and new tests pin that every rule is selectable and prints the same alone as in the full run. The harness module went from 550 s to 166 s locally (child CPU -78%); the PR's CI harness job took 166 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-11-14 # Conflicts: # .github/workflows/ci.yml # README.md # ci/proofs/platform-generated-enums.json # ci/proofs/serde-json-immutable.json # ci/readonly_dependencies.py # ci/readonly_generated.py
Upstream's owned-presentation refactor removed plx_machine::idle. A favourite's server answer and the socket's Home refetch land on workers, so they wake through present::wake_from_worker; the press itself already presents its own frame.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Merges PlxNative
mainup to 2e1b78d (plxnative#1386), 19 commits after plxnative#1364. No upstream release is pending, so this is not release-aligned;PLXNATIVE_BASEstays0.11.0-dev.Conflicts resolved:
.github/workflows/ci.yml: upstream'sruns-on: ubuntu-latest, FinNative's 15-minute harness limit kept.README.md: FinNative's README kept (upstream's new "Also for LG webOS" section is PlxNative's own).ci/proofs/*and the tworeadonly_*.pycertificate hashes: regenerated for the merged manifests and lock; FinNative'si18n/mod.rshash put back inI18N_SOURCES.Upstream bumps the Rust toolchain to 1.100.0 (#6f9bbc9e).
Generated by Claude Code