fix: JJ Bookmarks on merges, speed k-boundary, and time cursor on timers - #688

What

• JJ Bookmarks lists bookmarks on different heads separately: 🔖 feature-a, feature-b instead of 🔖 feature-afeature-b when @ sits on a merge of two bookmarked heads.
• Input, Output and Total Speed switch to k where the value would round up to 1000: 999.96 t/s shows 1.0k t/s instead of 1000.0 t/s.
• Block Timer, Block Reset Timer and Extra Usage Utilization no longer say "time cursor" in the editor after the widget picker turns a usage bar with the time cursor on into one of them. They never draw a cursor and have no key to turn it off.

Why

• JJ Bookmarks runs jj log --no-graph -r 'heads(::@ & bookmarks())' --template bookmarks. jj log prints each commit's template output back to back, with nothing in between, so two heads give feature-afeature-b. The widget splits on whitespace, which only separates bookmarks on the same commit.
• Speed: formatSpeed compared against exactly 1000, then rendered with one decimal, so 999.95–999.99 rounded up inside the "t/s" range: formatSpeed(999.96) returned 1000.0 t/s. The same happens at any precision: with whole numbers, 999.6 rendered as 1000 t/s. formatTokens already moves up a unit at the rounding boundary (999,950 tokens is 1.0M, not 1000.0k).
• Time cursor: the cursor is the cursor metadata flag of the usage bars (Session Usage, Weekly Usage and the per-model weekly widgets). Changing a widget's type in the picker keeps its metadata, and the shared modifier text listed "time cursor" for any widget in a bar mode with the flag set. So after turning a Session Usage long bar with the cursor on into a Block Timer, the editor row read Block Timer (long bar, time cursor), with no cursor drawn and no (t) key to clear it.

How

• JJ Bookmarks: the template is now bookmarks ++ "\n", so each head's output ends with a newline. The whitespace split then works across heads as it already did within one, and the output for a single head is unchanged.
• Speed: formatSpeed switches to k at 1000 - 0.5 / 10 ** decimals, using the effective decimals of the widget's number format: 999.95 by default, 999.5 for whole numbers, 999.995 at two decimals. This is the same boundary formatTokens uses for M. Values below the boundary and all k values render as before. There's no M tier for speeds, so values from 999,950 t/s up still read 1000.0k t/s and above, as before.
• Time cursor: the modifier text now takes an includeCursor option, like includeCompact and includeDate, and only the usage percent widgets pass it, since they are the ones that draw the cursor. The metadata itself is left alone, so changing back to a usage bar keeps the cursor setting. The editor row of the usage bars is unchanged.

Demo

The status line piped from a sample transcript (99,996 output and 99,994 input tokens over 100 s) with a stub jj standing in for a repo where @ is a merge of two bookmarked heads. jj isn't installed here; the stub prints what jj log --no-graph prints for each template, one head's output after the other. Input Speed (999.9 t/s) and Total Speed (2.0k t/s) are unchanged.

Powerline: before

Powerline: after

Plain: before

Plain: after

In the TUI, a Session Usage long bar with the time cursor on is changed to Block Timer in the type picker (sample preview data):

Powerline: before

Powerline: after

Plain: before

Plain: after

Testing

• New tests, each failing on main first: JJ Bookmarks on two heads (the template and the parsed output); speeds at the boundary by default (999.96, 999.99 vs 999.94), with whole numbers (999.6 vs 999.4) and at two decimals (999.996 vs 999.994); the modifier text of Block Timer, Block Reset Timer and Extra Usage Utilization with carried-over cursor metadata. The existing test for the usage bars' "time cursor" modifier still passes.
• bun test: 2788 pass, 0 fail. bun run lint passes.
• The changed test files under Node (Vitest): 108 pass (102 on main).
• Built CLI under Bun 1.4.2 and Node 26.10.0 with the speed and JJ Bookmarks line above, the sample transcript and the stub jj: all runtimes give identical output in plain and Powerline modes, and it differs from main only in Output Speed and JJ Bookmarks. With a Session Usage bar with the cursor on, and with the baseline line, the output matches main. The TUI opens and exits cleanly under both.