# Two-column layout: replace calc()-based heights with a viewport-locked flex shell

## Background

musehub#84 fixed the original bug (sticky sidebars clipped and couldn't scroll)
by moving from `position: sticky` to a bounded-height pattern: each column got
`max-height: calc(100dvh - var(--sticky-offset, var(--header-height)) - ...)`
plus `overflow-y: auto`.

That pattern has now broken twice, both for the same underlying reason —
`--sticky-offset` is a **static CSS guess**:

```scss
body:has(.repo-tabs) {
  --sticky-offset: calc(var(--header-height) + var(--repo-tabs-height));
}
```

It only accounts for the navbar + repo-tabs bar. Any page with additional
chrome between the header and the two-column grid (a toolbar, filter bar,
hero section, stat strip) has a wrong `--sticky-offset`, so every column's
computed `max-height` is off by however tall that extra chrome is:

1. **Footer-overlap regression** (fixed this cycle) — the calc() put on the
   *grid wrapper itself* (an earlier iteration of this pattern) claimed a
   full viewport's height measured from wherever the grid happened to sit,
   pushing the footer down to overlap still-visible content on pages with a
   hero/eyebrow above the grid.
2. **Unreachable pagination regression** (this ticket) — on the issues list
   page, `.isl-main`'s `max-height` doesn't account for the issues-list
   toolbar above `.isl-layout`, so the column's box extends past the
   visible viewport with no way to scroll to the pagination controls at the
   bottom.

Patching the formula a third time (adding yet another term to
`--sticky-offset`, or a per-page override) just sets up the next page that
adds chrome above its grid to hit the same bug again. The pattern itself —
hand-deriving "how much vertical space is left" per page in CSS — is the
defect, not any single formula.

## Goal

Replace the calc()-based height budget with a **viewport-locked flex shell**,
the same pattern Gmail/Slack/VS Code/Notion use for independently-scrollable
sibling panes: the shell claims `height: 100dvh` once, every fixed-height
sibling above the scrollable region is `flex-shrink: 0`, and the scrollable
region itself is `flex: 1; min-height: 0`. The browser computes "whatever's
left" automatically — no calc(), no magic numbers, no per-page offset to get
wrong, and this bug class becomes structurally impossible rather than
something to keep re-patching.

Confirmed with gabriel: Gmail's own layout (header fixed, no footer while
inside the app view, two panes each scroll independently within a
viewport-locked shell) is the reference UX to match.

## Design

- `base.html` gets an opt-in body class, `body-app-shell`, set only by
  templates that use a two-column layout. Pages without one (home, plain
  forms, docs/marketing pages) are untouched — normal document scroll,
  footer visible.
- `body.body-app-shell`: `height: 100dvh; overflow: hidden; display: flex;
  flex-direction: column;`. Navbar and repo-tabs: `flex-shrink: 0`.
  `#content`: `flex: 1; min-height: 0; overflow: hidden; display: flex;
  flex-direction: column;`.
- The footer is hidden on `body-app-shell` pages — matches Gmail; a
  marketing-site footer doesn't belong on a dense two-column app screen.
  Pages that keep normal scroll keep the footer exactly as today.
- Every wrapper between `#content` and the actual two-column grid (per-page
  toolbars, etc.) becomes `flex: 1; min-height: 0; display: flex;
  flex-direction: column;` so the "fill remaining space" chain isn't broken
  partway down.
- Each column (`*-main`, `*-sidebar`) keeps `min-height: 0; overflow-y: auto;`
  — no `max-height`, no `calc()`.
- `--sticky-offset` / `--repo-tabs-height` are deleted once nothing
  references them.

## Affected pages (11) — all resolved

- [x] Repo home (`_repo-home.scss` / `repo_home.html`)
- [x] Issues list (`_issues.scss` isl-* / `issue_list.html`) — the page
      that was actually broken; also where the two real gotchas below
      were found, and where the zoom-clipping bug (see Findings) was
      root-caused.
- [x] Issue detail (`_issues.scss` isd-* / `issue_detail.html`)
- [x] Proposal detail (`_proposal-detail.scss` / `proposal_detail.html`)
- [x] Explore/browse (`_explore.scss` / `explore.html`) — also fixed an
      unrelated pre-existing bug found along the way: `container_extra_class`
      was missing its leading space, silently breaking `.container-wide`
      matching and `.page-container`'s padding-strip on this page.
- [x] MCP docs (`_docs.scss` / `mcp_docs.html`)
- [x] Muse dev docs (`_muse-docs.scss`, 17 `docs_muse_*.html` templates,
      one shared CSS file) — landing page (`docs_muse.html`) correctly
      left untouched, single-column normal scroll.
- [x] Blame (`_blame.scss` / `blame.html`)
- [x] Blob/outline panel (`_blob.scss`) — confirmed no edits needed; this
      page self-caps the panel's own height directly against the viewport
      rather than matching a sibling column, so it works fine within
      normal document scroll. Deliberately not converted to app-shell.
- [x] Profile graph/DAG sidebar (`_profile.scss` `.graph-sidebar`) — turned
      out to have **no template anywhere in the repo**; not a page to
      convert. Left in place with a TODO (likely scaffolded ahead of the
      code-domain symbol-graph visualization, per gabriel) instead of
      deleting — re-check when that work starts.
- [x] `_layout.scss`'s shared `.layout-two-col`/`.layout-sidebar` utility
      classes — audited, confirmed **zero templates ever used them**.
      Deleted the base rules, the dead responsive breakpoint overrides,
      and `layout-sidebar.test.ts` (existed solely to guard these
      selectors — nothing left to test once they're gone).

Every page verified individually via headless Chromium (Playwright) driving
the real rendered page with genuine `mouse.wheel()` events — not just
reading CSS — before moving to the next. `--sticky-offset`/`--repo-tabs-height`
are still referenced by the (now zero) non-app-shell two-column pages, so
left in place rather than deleted; nothing currently depends on removing them.

## Coordination note

aaronrene picked up musehub#129 (rich social preview cards / OG images) at
the same time this work started. That ticket touches
`musehub/api/routes/musehub/_ui_helpers.py`, `ui_repo.py` (backend route,
OG tag wiring), and a new `/opengraph-image` endpoint — no overlap expected
with this ticket's scope (base.html, SCSS, page templates' layout markup).
Flagging here so neither of us has to guess.

## Findings while converting the issues list page (first page done)

Two real pitfalls found via headless-browser verification (Playwright,
not just reading CSS) — both are the actual explanations for scroll
failures hit during this work, not the zoom theory originally suspected:

1. **`align-items: start` on the grid.** Carried over from the old
   max-height-per-column pattern, where each column set its own height
   independently and `start` kept the shorter column visually anchored to
   the top. In the flex-fill model the grid row's height comes from the
   grid container's own (flex-derived) height — `align-items: start` makes
   grid *items* size to their own content instead of stretching to fill
   that row, so a column's `overflow-y: auto` never gets a bounded box to
   clip against. Must be `align-items: stretch` (or omitted — that's the
   default) on any grid converted to this pattern.
2. **A `.card` (or similar) child with its own `overflow: hidden`, inside
   a `flex-direction: column` scroll container.** Flex items default to
   `flex-shrink: 1`. When content exceeds the scroll container's bounded
   height, the browser shrinks the child to fit *instead of* letting it
   overflow — and if that child has its own `overflow: hidden` (often
   there just to clip rounded corners, as in `issue_list.html`'s
   `<div class="card" style="overflow:hidden">`), the excess content is
   silently eaten with no scrollbar anywhere. The parent's own
   `overflow-y: auto` never triggers because nothing actually overflows
   *its* box. Fix: `flex-shrink: 0` on that child.

Neither of these is specific to zoom or `--sticky-offset` — they're
generic flex/grid gotchas that would bite any page using this pattern.
Checking for both on every remaining page before considering it done.

Verified via headless Chromium (Playwright) driving the actual rendered
page and reading `scrollHeight`/`clientHeight`/`scrollTop`, not just
reasoning about the CSS — recommended for verifying the remaining pages
too, since two prior "should work" theories (the `--sticky-offset` gap,
then the zoom/dvh interaction) both turned out to be wrong or incomplete
without it.

- [x] Issues list (`_issues.scss` isl-* / `issue_list.html`) — verified:
      independent scroll on both columns at 700px viewport, mobile
      breakpoint (700px width) stacks to single-column normal scroll with
      footer visible, desktop footer hidden and not overlapping.

## Third finding — zoom clips the last ~20% of every app-shell page

A third, more fundamental bug surfaced after the two gotchas above were
fixed and the issues-list pagination link was *still* unreachable:
`zoom` scales an element's own rendered box, not just its descendants.
`body.app-shell { height: 100dvh }` zoomed 1.25x paints ~125dvh tall — the
25% below the real viewport was silently clipped by `overflow: hidden`,
no matter how far any descendant scrolled internally. This is what was
actually causing "can't reach pagination" from the very start; the
align-items/flex-shrink bugs were real but separate, and had been masking
this one underneath.

Fixed with a shared `--zoom-factor` custom property (used by both
`body`'s `zoom` and `body.app-shell`'s `height: calc(100dvh /
var(--zoom-factor))` compensation), plus overriding a competing
`body { min-height: 100dvh }` from `_components.scss` that was clamping
the compensated height back up (`min-height` wins over a smaller
`height`).

Confirmed working live by gabriel on the issues list page after this fix,
with a real "Next →" pagination link (41 real + test issues) reachable
via genuine mouse-wheel scroll.

## Status: done

All 11 checklist items resolved, each verified via headless Chromium with
real `mouse.wheel()` input (not just DOM `scrollTo`, which bypasses hit
testing and would have masked the zoom-clipping bug). Local test issues
created for verification should be cleaned up before closing. Branch
`fix/two-column-app-shell` not yet merged to `dev` — pending final
review/merge decision.