feat(app): nest forked sessions under their parent in the session list

[chphch]

Forked sessions already persist parentSessionId in their (encrypted) metadata — it's set on every fork path (the Fork quick-action, duplicate-from-message, and the MCP open_session tool) and already surfaced as the "Forked From" link on the session info screen — but the session list ignored it, so a fork showed up as an unrelated row sorted only by recency and you couldn't tell at a glance what was branched from what.

This renders forked children directly under their parent within each list section (the active project group, and each by-date group), indented by fork depth with a └ tree connector. A child whose parent isn't in the same section falls back to depth 0 so nesting never crosses section boundaries, and the visual indent caps at a max depth so deep chains never march off-screen (the underlying depth stays accurate).

Gated behind an experimental expForkNesting setting, off by default. With it off the session list is byte-for-byte the upstream layout — rows keep their original order and stay at depth 0, so the renderer draws no indentation and no connectors. Users opt in from Settings → Features.

Shape of the change

• utils/forkLineage.ts (new) — the pure reorder + indent math, dependency-free and unit-tested.
• sync/storage.ts — SessionRowData gains parentSessionId and forkDepth; a new buildOrderedSessionRows applies the lineage ordering when the flag is on and returns rows untouched at depth 0 when it's off, so the flag check lives in exactly one place.
• Renderers — both the compact active-session rows and the full by-date rows consume forkDepth for their indent padding.

Proof

Synthetic parent → two children → grandchild sessions rendered through the real SessionsList on Expo web (standalone server, dev_token URL bypass, driven with Playwright). Left = flag off (forks scattered by recency, lineage invisible); right = flag on (children nested under their parent with a └ connector; the grandchild sits one level deeper). Both render paths are shown — the compact active group (top) and the full by-date "Today" group (bottom).

Verification

pnpm typecheck clean; app suite green. utils/forkLineage.test.ts covers the ordering — multi-level chains, multiple children of one parent, the parent-not-in-section fallback, a parent-cycle guard, and deep-equal reference stability.

[chphch]

Rebased onto current main — this was CONFLICTING, and the rebase turned up that the branch had also become a no-op.

Why it stopped working

The session list moved to project cards. buildSessionListViewData now emits only projects-header / project items, so the helper that stamped forkDepth lost its last caller and every row rendered at depth 0 — the indent code in CompactSessionRow was alive the whole time, just never handed a non-zero depth.

What changed

• Nesting is applied in ProjectGroup.tsx, where the rows actually render, instead of where the list data is built. That also fixes a latent bug in the original approach: the list is filtered after it's built (archive toggle, search box), so a depth stamped at build time could leave a child indented under a parent that had been filtered off-screen. Stamping at render time makes that impossible.
• getSessionShortcutIdsInDisplayOrder nests the project branch too, so the ⌘1–9 palette numbering matches what's on screen.
• Removed the now-dead buildOrderedSessionRows and the orderForks parameter of buildSessionListViewData — with nesting done at render time the list no longer needs rebuilding when the setting flips; the component re-renders on its own.
• The test fixture now sets parentSessionId / forkDepth. The feature commit makes these required on SessionRowData, so without this tsc was red.

pnpm typecheck clean, 48 tests pass.

Proof

Three sessions in one project card, created via the daemon's /spawn-session with an explicit parentSessionId, so the middle one is a genuine fork of the row above it. Row identity was confirmed by clicking through to /session/<id> rather than inferred from position.

Setting on — the child is reordered directly under its parent, indented with a └ connector:

Setting off — original newest-first order, all three rows at the same offset, no connector:

Narrow width (420px) behaves identically: on / off.

[chphch]

Fixed a visibility bug in the connector itself, folded into the first commit.

The connector is drawn with theme.colors.divider; a selected row's background is theme.colors.surfaceSelected. On web those are the same value in light theme (#eaeaea), and near-identical in dark (#292929 vs #2C2C2E) — so the connector was invisible on exactly the row you are currently inside. Unselected rows sit on surface (#ffffff), which is why it looked fine until you opened the forked session.

ForkLineageConnector now takes selected and switches to theme.colors.textSecondary when set; both call sites pass it.

Measured rather than eyeballed, since this is a contrast bug — sampling pixels from the rendered app at 1280:

state	connector px	row background px	per-channel delta
selected, fixed	rgb(73,69,79)	rgb(234,234,234)	[161,165,155]
selected, pre-fix color forced	rgb(234,234,234)	rgb(234,234,234)	[0,0,0]
unselected (control)	rgb(234,234,234)	rgb(255,255,255)	[21,21,21]

Forcing the old colour back on the same element reproduces the bug exactly, and a whole-image diff between the two is 46 pixels — all inside the connector rect, nothing else moved. The unselected rendering is unchanged.

Scope of the measurement: light web theme and the compact row path. The SessionsList row path and dark theme take the identical one-line prop but were not exercised by pixels.