Skip to content

Merge upstream 2026-11-19 - #43

Merged
tiper merged 13 commits into
mainfrom
upstream-merge-2026-11-19
Nov 19, 2026
Merged

tiper merged 13 commits into
mainfrom
upstream-merge-2026-11-19

Conversation

@tiper

@tiper tiper commented Nov 19, 2026

Copy link
Copy Markdown
Owner

Merges PlxNative main up to e9924f0f (plxnative#1469), 9 commits, as a merge commit. No upstream release is pending, so this is not release-aligned; PLXNATIVE_BASE stays 0.11.0-dev.

Resolutions:

  • src/app/run.rs (Deliver typed text to any field with the keyboard, not only Search plxnative/plx-native#1428, typed text reaches any field with the keyboard): upstream's text_owner / text_panel_edge gate replaces FinNative's own text_entry gate, so the Jellyfin sign-in form receives text through the shared path (its OK raises the keyboard through Delivery::Keyboard, which system_keyboard_up tracks). FinNative keeps only one thing on top: a password field's text (TextEntry::Secret) is refused by the input recorder.
  • screens/search/mod.rs: upstream's new caption_policy field beside FinNative's raise_on_arrival.
  • ci/readonly_generated.py: upstream's file plus FinNative's i18n source hash; certificate regenerated.

Ledgers unchanged. check-machine-shape ok, ci/test_readonly_generated.py OK.


Generated by Claude Code

GLinnik21 and others added 13 commits November 18, 2026 17:21
…r a hero press (plxnative#1442)

Device and simulator scripts now reach elements by accessibility id instead of walking to them with direction keys, check where focus landed before taking a shot, and cannot press a title without naming it. This is the harness half of the series, after the ids (plxnative#1431) and the `focus:`/`press:`/`state:` tokens (plxnative#1432).

- **`tools/tv-session.sh` gains `focus <id>`, `press <id> [--rk RK]` and `state [--out FILE]`**, all under the TV lock.
  - Ids and ratingKeys are checked locally before anything reaches ssh or the FIFO.
  - A press on an element that acts on a title needs `--rk`, and the app refuses the press if that title is not the element's.
  - `key` refuses by-id words, so the verbs are the only way to send one.
  - The FIFO write escapes `'` for the remote shell and refuses any whitespace inside a token, because the app splits on whitespace.
  - `state` sends a random nonce and polls `plxnative-a11y.json` until the nonce matches.
- **`tools/a11y.py` is the simulator helper.**
  - It provides `focus`, `press(subject=)` and `expect_focus`.
  - Its command line applies the same title guard, with `--subject`. The library `press()` is unguarded, for scripts that drive the mock.
- **Every element whose press acts on a title it names now carries that title's ratingKey as its `A11y` subject.** It is read the way the activation reads it, with no state and no IO.
  - This covers the Home hero, the item menu, the Detail hero (a started show names the episode the press acts on), the Play Version rows, the player HUD's queue and restart controls, the failure screen's retries, the More menu's quality rungs, and the track and subtitle rows that write to the server.
  - `SUBJECT_BEARING` is the one list. A test holds it equal to every file that builds a play, queue-step or write message.
- **Home's hero/grid door.** A by-id press queues its focus move and its OK in the same frame, so it used to act on the element focused before the move.
  - A scripted focus move now clears that remap.
  - While a key's move has left the picture behind the focus, any by-id press of the focused element is refused as `unloaded` until the picture catches up.
  - A key move followed by a quick OK keeps today's behaviour.
- **The carousel holds still under a hero press.** While a press is armed on a hero button, the countdown restarts and a Flip or edge key is refused, so the commit acts on the title the press was accepted against. This also fixes a person's OK just before a turn.
- **A by-id press resolves against the live focus, and every delivery reads the live press state.** A by-id press batched after a key sees the focus that key moved, and a screen's tick sees a press armed in the same frame.
- **The Watchlist captures drive by id only.** Focus is checked through `state` before each shot, and any `a11y: refused` line fails the capture. `tests/player_shots.sh` presses the HUD by id.
- **The docs carry two rules:** prefer ids, walking only where focus travel is the subject; and the subject rule, pointing to `SUBJECT_BEARING`. They are in the `tv-session` and `ui-sim` skills, `docs/agent-reference.md`, `tests/README.md` and `rust-modules/ui/src/CLAUDE.md`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
plxnative#1459)

## Summary

The glass menus (Library and Watchlist Sort & Filter, the +N list, and the profile menu) are now frosted glass rather than a grey field: the frozen page behind the panel is blurred once, under a thin face of rgb(58,58,64) at .36 over rgb(40,40,46) at .44 (the owner's pick from the mock). The pass that follows makes the frost and the menus' existing close costs fit the TV's frame budget together.

**The frost** (`gfx::FrameCache`, profiler phase `glass.frost_build`; `glass.frost` stays the name of the experimental full-resolution frost quad in `glassload.rs`):
- Once per page snapshot, the region under the settled panel plus a 120 px margin (`FROST_MARGIN`, aligned to 8) is reduced 1/4 → 1/8 and blurred with two Kawase passes (d = 1.5, 2.5; σ ≈ 24 px). The panel composites that texture; nothing is blurred per frame.
- Legibility keeps every contrast floor. The body is capped at `FROST_BODY_CEILING` = 0.333, derived rather than tuned: the tertiary ink reaches 3:1 at a body of 0.3372, minus 1/255. A soft knee from 0.9 of the ceiling (`y = knee + d·o/(o+d)`) keeps a bright page from reading flat; at 0.6 the frost was visibly darker than the mock.
- The frost is a compile-time variant of the glass shader (`#define PLX_FROST`, `FS_GLASS_FROST`) linked with `link_optional_program`. If it fails to link, the log says `glass frost prog unavailable` and the menus keep the field face (`MENU_GLASS_FIELD_TOP/BOT`, .78/.86). The plain glass program is unchanged. The ~45 `static mut GL_*` uniform locations became a by-value `GlassUniforms`.

**The body bake** (`FrameCache::frost_bake_owed`, `fs_frost_bake.frag`, phase `glass.bake`): the modal dim and the ceiling are applied once into a 1/8-resolution texture, keyed by the frost epoch, the region, the dim (field texture, `tex_ledger::revision`, tint) and the cap, instead of per pixel on every frame. The dim reaches the bake as `ModalPresentation.dim` (`ModalStack::first_dim` → `UnderlayField::dim_of`, `Dispatcher::modal_presentation_at`). A page-sourced dim that latches after the head's bake wakes exactly one corrective frame (`DimSink::dim_settled` → `popover::host::frost_body_stale`).

## The account menu's close no longer re-takes the page

The account menu is a Sheet, which freezes its host while open, so `bind_owner` re-took the page on the close's first frame (14.9M GPU cycles, 133% of a refresh). Now a layer that has only started closing drops the ground stage and keeps the page snapshot for every presentation style (`ModalCacheKey::began_closing` returns whether it did; `kept_host_live` and `live_hosts` are gone).

That is safe only if a page that changes under a cover says so. Pages now report a setting that moved under a cover at the head of `step`, through one shared helper (`registry::moved(&mut held, now)` plus `fx.invalidate(Provenance::Nav)`): Home, Library (including `sort_value`), Recommended (`deck_press`), Collection, Person, Search, Detail and Watchlist (caption policy and `rev`). Home damages on `FocusMoved` only when the focus or the `place()` fingerprint moved, and on `RestoreMemory` only when `projected_generation` or `place()` changed.

## The morph is frosted the whole way, and its rows show earlier

The owner's direction from an iOS 26 recording: the droplet never reads as clear glass. `Frame::clear`, `clear_at` and `FROST_FROM`/`FROST_TO` are gone; the morph hands the glass only its join (clarity 0). From the first glass frame through the close's hand-back the body is the frost under `FROST_FACE`; a frame where the frost is not built yet wears the latched field under the heavier face, never the sharp page. `draw_shadowed` asks for the frost before the field check, so the build lands at the head of the open's second frame (O2) and stands when the glass first draws; nothing is added to O1. The rows fade in from `ROWS_FROM_AT` 0.05 (was 0.55; opaque at 0.8, was 0.96), first visible on the third glass frame (~33 ms) at a lens scale of ~0.2, and `content_alpha` is capped at `ground_alpha`. The close's row timing is unchanged.

## The opener is handed back the frame the droplet is gone

After the profile menu closed, the "Demo" chip sat lighter for ~300 ms and then snapped darker (the owner's TV clip). The droplet's close ends ~14 frames after Back, but the surface (and with it the opener's lift, which paints the chip over the lighter latched field) was retired only when the Sheet container's spring rested, ~17 frames later. `Surface::faded` now also retires a closing surface on `Screen::handed_back`, for every style. In a sim film of the same scene the chip patch read 86 for frames 186-203 and 77 after; with the fix it reads 77 from the first hand-back frame. Regression tests: `a_morph_menus_surface_and_lift_end_on_the_frame_its_droplet_is_gone` (fails without the fix) and the rewritten `account_menu_route_tests::the_morph_begins_and_ends_on_the_chip_as_the_bar_draws_it` (asserts the surface leaves at the hand-back and the bar then draws the chip at its resting rect; red without the fix).

## Tried and reverted: folding the dim into the page

The dim was folded into the held-page blit (`fs_dim_fold.frag`) to drop its separate full-screen pass. On the TV (same session, A-B-A, 11 cycles each) it COST GPU time on every dim-visible frame: Library open steady 9.36 / 9.34M with it and 9.05M without (+0.31M), account 7.60 / 7.60M against 7.19M (+0.41M), with 0.02M drift between the two fold-on runs. The `scrims` phase read 0 with the fold, but the folded draw landed in no phase, so only the frame total showed the cost. It is reverted in this branch.

## Measured on the TV

All builds `ARM_PROFILE=release`, debug flavour, mock accounts only (the synthetic PMS checked before any key; the account menu signs in through the loopback plex.tv stand-in). GPU_ACTIVE from Mali HWCNT, median per frame role over 11 cycles; the budget is 11.2M per 60 Hz refresh.

- **base:** main at 952d0a4 (plxnative#1444).
- **frost only:** the frost before this pass (run 3).
- **this branch:** the measured shape is run 6's fold-off build (ad49973 with the fold forced off; the branch's revert removes the same code). Run 5 measured ad49973 with the fold on; its pacing column is from that build.

| | base | frost only | this branch |
|---|---|---|---|
| Library, rows visible (100-200 ms) | 10.74M | 11.48M (26 frames over) | 10.40M |
| Library, open steady | 9.42M | 10.44M | 9.05M |
| Library, close C3 | 10.42M | 11.81M (11 over) | 10.29M |
| Library rows window, missed refreshes | 1 | 23 | 0 (run 5, fold on) |
| Account, close's re-take frame (C2) | 14.90M (133%) | 14.91M | 5.10M |
| Account close, missed refreshes | 25 | 33 | 1 (run 5, fold on) |
| Account, open steady | 7.27M | 7.63M | 7.19M |

- The frost's own build is one frame per snapshot (p50 0.68M; timed on the TV as `glass.frost`, renamed `glass.frost_build` afterwards) and the panel composite (`glass.frozen`) costs less than the base glass's (run 5: 4.19M against 4.63M).
- Every program linked on the device: no `glass frost prog unavailable`, `link FAILED` or `shader compile FAILED` line in any log.
- The DISPLAY captures of both menus show no banding, seam or flip.
- The base and frost-only columns are from run 3, the third from run 6: different sessions. Run 6's fold-on build reproduced run 5 within 0.05M, which bounds the drift between sessions.
- Not re-measured on the TV: the frosted-morph and hand-back changes (seen on the TV by the owner, not profiled), the pacing of the final shape (run 6 had no pacing leg; its frames are 0.3-0.4M cheaper than run 5's), and the commits after ad49973 (the revert, the merge of main, a test-isolation fix).

## Not fixed here

- O1, the open's capture-and-latch frame, is over budget on main too (12.2-12.9M); the second frame is late in 19-21 of 21 cycles. Pre-existing, unchanged here.
- The account menu's open missed 6 refreshes against 2-3 on base in run 5 (fold on). Without the fold its open frames are back at base (7.19M steady against 7.27M); its pacing was not re-measured.
- `glass.bake` records at 2.0M during the dim ramp; not attributed.
- The review noted that a close may rebuild the frost once (the close's first frame re-captures the page after the ground stage is dropped, which moves the page epoch). The run 5 HWCNT leg recorded one frost build per open-close cycle (11 over 11 cycles), and the close frames are under budget (Library C2 7.16M, account 5.50M), so it was left as is.
- Damage under a cover not yet narrowed: Collection and Person on Enter/Uncover, `cards/grid.rs:355` and `shelf.rs:337/359`, `library/status.rs:53`, the `others_up` path (no Uncover), and the container chrome (not audited).
)

Two doc comments still named `library::LibraryScreen`'s `pop_from_rest`, which no longer exists, as the read gated on `By::Dir | By::Pointer`. One is the doc on `FocusTarget::FirstInGroupAnimated` in `machine.rs`, the other the module doc of `ui/src/containers/tabs.rs`. Both now tell that bug's history in the past tense, under its old name. They add one sentence for today: the shared card grid reads a `By::Dir | By::Pointer | By::Script` move as deliberate (`ui/src/cards/mod.rs`'s `Seen::of`), with `By::Script` since plxnative#1432.

`machine.rs` is pinned by hash in `ci/proofs/platform-generated-enums.json`. Its new hash is appended to `MACHINE_SOURCES` and to the certificate's list, and `CERTIFICATE_SHA256` is recomputed. Every changed line in `machine.rs` is a `///` line, and no other certificate field moves.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onment, and cut fleet lanes there (plxnative#1443)

launchd inherits nothing from the installing shell, so the hourly
build-gc --auto always resolved $PLX_FLEET_DIR to ~/plx-fleet even on
machines that moved the fleet to a second volume: external lanes were
reaped only as registered worktrees, never through the fleet-dir
sweeps. Bake the explicitly-configured value into the generated plist
(unset keeps today's behavior and adds no env key), cover it with a
second installer test, route the fleet-plan recipe through
${PLX_FLEET_DIR:-$MAIN/.claude/worktrees} so whole lanes — build trees
still inside their checkout — default onto the external volume, and
fix the recipe's pre-move MAIN path.
…lxnative#1428)

Settings > Player HUD > Install from address could not be typed into. Every typed-text
ingress in app/run.rs was gated on bridge::search_owns_input, from when Search was the only
text field: the FIFO's txt:, the television keyboard's SDL_TEXTINPUT and a replayed "text".
On any other page the text went to textinput's queue, which nothing reads.

Dispatcher::system_keyboard_up() now reads what the dispatcher already owns: the keyboard is
up for the page or surface that raised it. bridge::text_owner is that, or Search on top, and
all three text sites ask it. A txt: for a page that is not Search drains SDL first (keys
before text), and replay does the same. The gate is read at ingest, so a txt: in the same
FIFO batch as the OK that raised the keyboard still finds no owner. The television's keyboard
always commits frames after it opened.

Only Search's commits carry the SystemKeyboard(true) adoption edge (bridge::text_panel_edge).
Any other page takes text only while our own keyboard flag is already up, so there the edge
could only re-raise a keyboard that a key in the same drain had closed, which left focus
travel and Press delivery off.

Hostsim tests:
- typed_text_reaches_the_player_hud_address_field_and_its_fetch, red against main's gate;
- text_after_the_key_that_closes_the_hud_field_does_not_re_raise_the_keyboard.

On the dev television (debug build, mock guest), three things were verified: typing an
address and fetching it, the HUD-folder row being inert while that fetch ran, and Search
still typing. Typing on the LG on-screen keyboard itself was not determined.

docs/hud-plugins.md and docs/search.md now say who receives typed text, and what a text
record's panel flag means. The key_play / Pointer row of
docs/implicit-machines-audit-2026-11-16.md now says that whether a viewer can reach the PLAY
entry is not determined.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lxnative#1464)

A linked collection heading could remember a slot from before its cards loaded. Once Home pruned the slot's key, Down reused it and frame reconciliation moved focus to the top of the page.

Resolve the remembered key before publishing the move. Keep a recovered position within its destination entry and shelf. If the old position can no longer be resolved there, select the head of that shelf.

The regression covers late card arrival, pruning, and repeated re-entry. It verifies the immediate directional landing, replacement of the remembered key, and the absence of a second focus move.

Closes plxnative#1458.
…ubscription (plxnative#1468)

Adds .github/workflows/claude.yml. An `@claude` mention in an issue, a pull request
comment or a review runs anthropics/claude-code-action (pinned to v1.0.283), signed in
with the maintainer's Claude subscription through the CLAUDE_CODE_OAUTH_TOKEN secret. Its
main job is reviewing pull requests, outside contributors' forks included.

Only owners, members and collaborators start a runner; the action re-checks write access
and refuses bots. On a fork's pull request only a conversation comment works, because
GitHub gives a fork's review events no secrets.

On a pull request the action checks out the head and restores only .claude/, CLAUDE.md
and AGENTS.md from main, so skills, agents (through the .claude -> .agents links), nested
guides and tools/ are the contributor's. The job therefore:

- uses its own GITHUB_TOKEN (contents: read, no id-token), which cannot push a branch or
  a v* tag and cannot satisfy the Anthropic federation rule issue-labels.yml uses;
- turns off project hooks, skills, subagents and skill shell commands, and denies the
  tools/ scripts;
- denies curl, wget, gh, git push, git tag and Read of /proc (prefix rules; the token
  scope and the scrub are the real limits);
- strips Anthropic credentials from Claude's commands and sandboxes them with bubblewrap,
  installed through the shared bounded apt-install action.

Claude posts inline review comments through the action's tool. On issues it can comment
but cannot push a branch.
…-11-19

# Conflicts:
#	ci/readonly_generated.py
#	rust-modules/screens/src/search/mod.rs
#	rust-modules/src/app/run.rs
Upstream's plxnative#1442 makes every hero control that writes to a title name it
for the harness's press guard; the favourite toggle writes the page's item.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JLzd9SJCCZ5zFGv23XMA8h
Keeps upstream's ingest_text signature, which its hostsim tests call.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JLzd9SJCCZ5zFGv23XMA8h
The favourite writes to the server, so its two hero ids join the
subject-bearing list the harness guard and tv-session.sh check.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JLzd9SJCCZ5zFGv23XMA8h
@tiper
tiper merged commit 740ce80 into main Nov 19, 2026
12 checks passed
@tiper
tiper deleted the upstream-merge-2026-11-19 branch November 19, 2026 06:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants