Replay 0.9.1

This release is withdrawn. Its download would not open.
Bundle.module traps when it cannot find its resource bundle, and the release build
shaped that bundle differently from every local build, so the app died on launch. The zip
has been removed. Use v0.9.2, which fixes it and adds a check that runs the app before
publishing it.

---

Download

Replay-0.9.1.zip — unzip it and drag Replay.app to Applications.

macOS will refuse to open it the first time. Here is why, and what to do.

Replay has no Apple Developer ID yet, so macOS marks it as unverified and shows
"Apple could not verify Replay is free of malware" — offering only Move to
Trash or Cancel. That message is about a missing certificate, not about
anything found in the app. Every unsigned app gets it.

To open it:

1. Try to open Replay once and dismiss the warning.
2. Go to System Settings ▸ Privacy & Security, scroll down, click Open Anyway.
3. Confirm. It opens normally from then on.

Or one line in a terminal, which does the same thing:

xattr -dr com.apple.quarantine /Applications/Replay.app

Right-click ▸ Open does not work. That bypass existed for years and Apple
removed it in macOS 15. It is still the advice in most projects' READMEs.

Or skip all of that

Homebrew compiles it on your machine, so it is never downloaded, never quarantined,
and opens with no warning at all:

brew tap nurkamol/tap
brew install nurkamol/tap/replay-app    # the application
brew install nurkamol/tap/replay        # the command-line reader

A signed, notarised disk image — no warnings, nothing to click through — appears
here the day there is a certificate. The workflow that builds it is already written.

---

Everything below happened after 0.9.0 was tagged the same day. Two of them are the reason to
take this build rather than that one: a data-loss bug, and the last functional gap against
the reference closing.

Added

• PDF export, which the ledger had recorded as a decision rather than an omission after
  three WebKit routes died. The fourth route is not WebKit: ImageRenderer returns a
  CGContext, a PDF context is a CGContext, and a SwiftUI view draws straight into a page
  with no browser and nothing to paginate. Every earlier failure was a pagination failure, and
  the reference caps its own PDF at one page — so the honest shape is a summary that counts
  the whole span and names what it left off, rather than the HTML report again.
• A menu bar popover in place of the menu. The menu answered two questions well and could
  not answer more, because a menu is rows of text. This adds the day's total, the focus goal
  as a bar, the last three sessions with their applications' icons, and a pause control.
• tools/screenshots.sh — every surface captured in one command, about two minutes. Not a
  UI test and it makes no assertions: 951 contract checks had never caught a layout bug, and
  the bottleneck was never whether to look but that looking cost a dozen commands each time.
• Localisation, to the point where it needs a translator rather than a task. Every string
  a reader sees goes through Loc, whose key is the English text — so a missing translation
  falls back to correct English rather than a raw identifier, and the contract keeps comparing
  the strings it always did. tools/strings-audit.mjs holds it there.
• A landing page at nurkamol.github.io/replay-swift, and a v0.9.0 release with a
  zipped app, since Homebrew and a source build are not routes everyone has.

Fixed

• Two copies of Replay could zero each other's live session. ActivityStore.open() closes
  any session left with no end — right after a crash, destructive while another copy is
  running, since it sets ended_at = started_at, duration = 0 on the first instance's
  in-flight session. Both then wrote to one SQLite file with no busy timeout, so a contended
  write returned SQLITE_BUSY and the event was lost with no error anywhere. A second launch
  now hands the front to the first and exits; the store waits five seconds for a lock. Found
  from a question, not from a check — whether running the Glaze app alongside this one could
  explain a crash. It could not. This was underneath the question.
• The menu bar panel opened mid-sentence. The popover is reused so the button can toggle
  it and remembered its last size, while the content changes height — so taller content was
  clipped off the top, taking the header and the day's total with it.
• The year heatmap's weekday key sat a hundred points from the year it labelled. The grid
  anchors trailing so it opens on today, which is right; applied unconditionally it also shoves
  a grid that fits against the right edge. True at every window from about 1000pt, which is
  every default window, and found by the first light-mode render in the project's history.
• The week's rhythm strip left a dead gap before the durations at any wide window. Its
  560pt cap was redundant — the page measure already bounds the row — and was concealing that
  the strip is twenty-four bars that each want to fill, so removing the cap pushed the duration
  column off the edge entirely until it was given layout priority.
• A page never grew with its window: every surface was capped at the prose measure and
  pinned left, so widening the window grew a dead column against the scroll bar.
• The Canvas field is still. Its sway was this port's own, cost a permanent 30fps redraw,
  and diverged from a reference whose graph does not move.

Known gaps

• Still not signed. The download is a zip and macOS will refuse it once; the release notes
  say why and how. Homebrew and a source build avoid it entirely.
• One language. The mechanism is built and every string goes through it; nobody has
  translated anything. Loc.count has two plural forms, which is English — Russian needs
  three, Arabic six. A .stringsdict is the seam waiting for a language to be chosen.