Skip to content

Version 1.3.47 - #798

Merged
maboa merged 20 commits into
mainfrom
release/1.3.47
Nov 12, 2026
Merged

maboa merged 20 commits into
mainfrom
release/1.3.47

Conversation

@maboa

@maboa maboa commented Nov 12, 2026

Copy link
Copy Markdown
Member

Release 1.3.47. Fixes #795, fixes #792, fixes #794, fixes #679, fixes #793, fixes #796, fixes #787, fixes #797.

Changed

maboa added 20 commits November 11, 2026 19:25
…light (#795)

Setting currentTime starts an asynchronous seek, and the media element keeps
playing the old position until it lands: about 18 ms from a local server, but
96 to 196 ms from a custom URL scheme in a WKWebView. Speech starts about
50 ms into a struck word's span, so the word's first syllable was heard.

The three skip sites in checkStrikeThrus now go through one helper that mutes
the player for the seek and gives the user's own mute state back on seeked,
or after 1.5 s if it never fires. A flag on the element keeps the playbar's
volume reflector from flashing the mute icon on every skip.

First change after 1.3.46: version markers and the two changed scripts' cache
stamps move to 1.3.47, release date cleared.
Deleting the first word after "[Mark] " also took the label's trailing space
and left the caret at the label's end, so the retyped word went into the
label span. The sanitise pass split the label back out, but the leftover
kept the parent's speaker class and zero duration: bold, untimed, hidden
with Speakers off, and the original word span was left holding a space.

When the span is a label already and its text starts with the label, the
label now stays where it is with its space restored, and the typed text
moves into the next word span, in the reported case the emptied original
with its timing intact. The caret follows the moved text so typing carries
on in the right span.
Testers deleted words expecting the audio to follow, and never found the
strike-out button, whose tooltip only named the formatting.

The button now says what it does: "Cut from audio and video (strike out,
⌘⇧X)" when words are selected, "Select words to cut from audio and video"
when they are not. The greyed button keeps the pointer so that second
tooltip can show (#679). ⌘⇧X / Ctrl+Shift+X strikes or unstrikes the
selection from the keyboard, and leaves text fields alone.

The first keyboard deletion of a whole timed word in a visit gets a one-time
dialog, "Deleted from the text, not the media", explaining the difference and
naming the shortcut. The deletion itself goes ahead. "Don't tell me again"
is remembered (noStrikeHint, cleared with the other dismissals in Settings).

The tour card is retitled "Cut words" to match.

Fixes #794, fixes #679
…e player reports time behind the sound (#795)

Muting for the skip's seek was not enough in Safari. WebKit's media element
reports currentTime about 0.27 s behind the audio it is producing: a tone
ladder tapped through an AnalyserNode changed tone 0.27 s before currentTime
reached the change, consistently, where Chromium was within 0.02 s. An ear
test on a real recording picked 0.26 s. So by the time the skip check saw
the cut, the ear was already a quarter of a second into the struck word.

In WebKit the mute now goes on 0.26 s before the cut, and the seek stays at
the cut, so no kept audio is skipped and, the sound really being at the cut
by then, nothing of the word before it is lost. Chromium gets no lead. The
mute and the seek are separate steps, and when the mute went on during the
lead and a seek then arrives by hand, the unmute waits for the skip's own
seek.

The spec parks the playhead 150 ms before the cut in WebKit and Chromium:
muted with the flag set in the first, untouched in the second, not yet
seeking in either, then the skip and the restore at the cut in both.

Fixes #795
… starts from it (#793)

The open and import paths signal the document's identity and undo history
takes a fresh baseline on the signal. The intro seed did not, so history
kept the baseline it took from the markup while the page was still parsing.
With the stock markup the two agree. A host that swaps in its own intro
before the seed runs ended up with a history whose oldest state was the
markup's intro: undoing past the first edit put that on screen over the
host's media, the host's intro as it was before the first edit was not in
history at all, and the autosave kept whatever the user ended on. A tester
undid one edit and landed on a transcript that did not match the video.

The seed now signals identity on success, as the other births do. The spec
plays the host, swapping the transcript at DOMContentLoaded, edits once,
undoes, and expects the host's words back and none of the markup's. Without
the signal it fails on exactly that.

Fixes #793
…log (#796)

The tour was fourteen cards. Five of them only pointed inside a dialog the
card before had opened, and the first two were both about NEW, before the
viewer had touched anything. In user testing people started by editing the
intro, which is what the first card told them to do, and the two buttons
they misread, cut words and skip the gaps, arrived fourth and fifth. One
tester said the tour was too long.

Now eight: the transcript, cut words, skip the gaps, transcript and
captions, export media, transcribe, recents, settings. What a dialog offers
is said on its button's card. The code that opened and closed dialogs, the
captions view and the Recents menu for a card is gone, with the CSS that
stopped dialogs fading while the tour ran. The transcribe card keeps its
neutral wording about where transcription happens, since a host that hides
some engines shows the same card.

The spec walks the eight cards, checks each sits on its own button and that
nothing is opened on the way.

Fixes #796
… button (#794)

The hint read the selection to see whether a deletion removed a whole timed
word, so a word deleted from a collapsed caret with ⌥⌫ (deleteWordBackward)
never counted: the selection is empty at that moment. It now asks the event
for its target ranges, which is what the browser is about to remove, and
falls back to the selection where a browser has none.

The dialog's sentence now carries the toolbar button itself, a copy of its
icon in a small bordered box, so the picture in the dialog is the button to
look for: "strike them out by selecting them and using the [button] button
or ⌘⇧X". For that the shared dialog accepts a built node as its message as
well as text. It still takes no markup.

Part of #794
…le a transcript is busy (#787)

In the captions view #hypertranscript is off the page, and engines paint
their loader and their result into it. Every engine clicked back to the
transcript view as its first step, but nothing required it: an engine that
followed the written contract without the click got no loader, since
showTranscriptNotice and setTranscriptBusy return early with no transcript
on the page, and threw on completion. And nothing stopped a switch to
captions while a transcription ran on screen, after which the result had
nowhere to land.

showTranscriptNotice and setTranscriptBusy(true) now leave the captions view
themselves, through one helper in editor-core, and the project library's
wrapper calls it before it reads the loader. While the transcript is busy
the captions segment of the view switch greys and declines the click, read
off the page's own state, since the busy attribute is also cleared outside
setTranscriptBusy when the user switches away. The lock lifts with that
switch, and a result finishing then goes to Recents as before. The engine
docs drop the step from the contract and the checklist. The engines keep
their own click, now harmless.

The spec plays a contract-only engine: started from the captions view, with
a captions click attempted mid-run, and finishing after the user has moved
to another project's captions.

Fixes #787
People click a word in the transcript expecting the player to go there, as
it does in the player and in most other transcript editors. A single click
only placed the caret. A double-click moved the playhead, and the Playback
setting from #541 made it play from there too, but nobody in user testing
found the double-click.

A click on a word now moves the playhead there while the media is paused,
highlights the word, and never starts playback. While the media plays a
click only places the caret, as before, so editing behind the playhead is
untouched. A drag that ends on a word is a selection, not a click, and the
second click of a double-click is left to the double-click behaviour. It is
a toggle on the Playback tab, "Click a word to move the playhead there", on
by default, because of the pause-and-fix habit: pause, click a word a few
seconds back, fix it, play, which then resumes from the fixed word. The
listener lives in the editor, not the vendored library. The tour's first
card says click rather than double-click.

The spec covers a paused click, a click while playing, the setting off
through the dialog and across a reload, and a drag selection.

Fixes #797
… the intro seed sends one (#793)

Three tests assumed the generation was zero after first load. Since #793 the intro seed signals identity on first boot, so they are off by one. They now read the generation first and expect one more, or the same, as the test says.
@maboa
maboa merged commit b8976ce into main Nov 12, 2026
1 check failed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment