Release tooling, CI gates, MRU sort, and a list scrollbar

[dsaad68]

Four changes, mostly independent. The first three came out of reading
alchemmist/lazy-tmux for ideas worth
borrowing; the fourth is a bug that showed up while testing them.

Release: goreleaser

.goreleaser.yaml replaces the hand-rolled for target in … loop in
release.yml. Same four platforms, but now tar.gz archives (carrying
kaku-tab.tmux and docs/ alongside the binary), a checksums file, a generated
changelog, and a Homebrew cask.

The binary keeps its plain name inside the archive — one unpacking to
kaku-tab_darwin_arm64 would install something kaku-tab.tmux's PATH lookup
cannot find.

A cask rather than a formula: goreleaser deprecated brews outright
(goreleaser check fails on it), and a cask also clears the macOS quarantine
attribute.

Before merging or tagging, publishing the cask needs two things that do not
exist yet:

1. a dsaad68/homebrew-tap repo
2. a HOMEBREW_TAP_TOKEN secret — a PAT with write access to it

Until both exist the release still succeeds and skips the cask (guarded on
skip_upload, tested with the variable unset and empty). The README's new
Homebrew section is wrong until the tap is created.

CI: lint and coverage gates

.golangci.yml, 11 linters, deliberately without revive's exported rule —
it produced 30 findings demanding doc comments on thin tmux wrappers and
bubbletea interface methods, which fights how this tree comments (why, not
what). The 5 real findings are fixed: two unchecked errors in popup(), an
ST1022 on DefaultTitleFormat, and dead minInt.

.testcoverage.yml excludes cmd/, internal/tmux, internal/kaku —
process wrappers whose unit tests could only assert arguments back at
themselves. Thresholds sit just under today's numbers so they ratchet rather
than block: total 40 (at 42.9), resolve 85 (at 90.8), ui 40 (at 44.7),
mru 80 (at 88.5). Each override regex was verified to actually bind.

CI gains lint, coverage, and goreleaser-check jobs. setup-go now reads
go.mod instead of a pinned 1.22 that no longer matches it.

Picker: @kaku-tab-sort

Takes tabs (default, unchanged), mru, or name.

The interesting part is that tmux cannot answer "where was I?"
#{window_activity} is output-driven, so a window running top is permanently
most-recent; #{session_last_attached} only moves on attach, which is exactly
what kaku-tab avoids doing when it retargets a tab. So internal/mru records
kaku-tab's own picks — into a tmux server option rather than a file, because
window ids mean nothing outside the server that issued them and a server option
lives exactly that long.

The window you are in is demoted one place. It always heads the history
(switching there is what recorded it), so leaving it at rank 0 would put the
cursor on a row whose Enter does nothing. pickHeaderWindow needed the same
treatment. Windows never picked keep index order behind the ranked ones, so a
fresh tmux server is identical to the old default.

Picker: scrollbar

A list running past the viewport looked exactly like one that ended there — the
rows just stopped at the frame. Both lists now carry a one-cell gutter with a
thumb sized to the fraction on screen.

Reserved always, blank when everything fits: a gutter appearing only on overflow
would re-budget every column the moment a query filtered a row away, and column
widths are a property of the table, not of the query. renderRow budgets
against a new rowWidth(); listWidth() keeps its old meaning for sizing the
preview.

Verification

• golangci-lint run — 0 issues
• go test -race ./... — pass; 18 new tests across internal/mru and internal/ui
• goreleaser check + a real --snapshot build: archives correct, injected version prints, cask renders
• Both sort modes and the scrollbar rendered against a live tmux server via the
  detached-session capture in CLAUDE.md

Left alone

go.mod says go 1.26.5 while the README says "Go 1.21+ to build". Pointing
setup-go at go.mod fixes CI either way, but which of the two is wrong is a
call about the project's support floor, not a CI detail.

🤖 Generated with Claude Code