fix(desktop): stop rendering an unhydrated project as lane headers with no rows

[thirstycrow]

What

Entering a project could render every branch/worktree lane with no session rows under it — permanently — while the drill-in read was returning the project's sessions the whole time.

Why it happens on current main

1. The overview tree ships lanes without rows on purpose. projects.tree builds with hydrate=False (hydrate=True is drill-in-only), and tui_gateway/project_tree.py:271-272 then empties every lane (group["sessions"] = []). Only projects.project_sessions returns rows. Verified against a real backend: the overview node for a project reports sessionCount: 43 with main → sessions: [], while the drill-in RPC returns the same lane with all 43 rows.
2. The sidebar falls back to that node while the drill-in read is pending (apps/desktop/src/app/chat/sidebar/index.tsx: enteredProjectTree?.id === overviewEnteredProject.id ? enteredProjectTree : overviewEnteredProject).
3. hasProjectContent counted the fallback as content — projectContent.sessionCount > 0 || repos.some(r => r.groups.length > 0) — and the overview node does carry sessionCount. So EnteredProjectContent rendered instead of the loading/empty state.
4. A lane with zero rows renders collapsed: workspace-group.tsx — defaultOpen = isProfileGroup || group.sessions.length > 0.

Net effect: the lane headers/tree render, nothing under them. And it can stay that way forever, because fetchProjectSessions resolved null for a superseded read exactly as it did for "no such project", and the hook committed that null as the project — so the empty fallback became the permanent state, with no error and no Retry affordance.

Reproduced end-to-end with real data: the drill-in payload + the app's own lane-merge helpers (mergeRepoWorktreeGroups, overlayLiveLanes) against a repo with 44 git worktree entries produce 16 lanes where the first (main) carries 43 rows — on the unhydrated fallback the same shape renders as 16 empty lanes.

Fix

• fetchProjectSessions reports a superseded read as ProjectSessionsSuperseded instead of resolving it. A stale failure is no more evidence than a stale answer, so it is reported the same way.
• The drill-in hook retries a superseded read (bounded: 3 × 150 ms, inside the existing effect) and reports failure once the budget is spent — which the sidebar already renders as its Retry affordance (SidebarLoadErrorState), instead of a project that looks empty.
• enteredProjectHasContent(project, hydrated) keeps the structure-only fallback from counting as content, so the skeleton/empty state renders instead of empty lanes. Rows placed by the live overlay still count as content, and a hydrated project whose rows were all filtered out (pinned) keeps rendering its real lanes.

Tests

npx vitest run src/app/chat/sidebar src/store → 2561 passed (209 files).
npx tsc -p . --noEmit, eslint, prettier --check on the touched files → clean.

• use-entered-project-sessions.test.ts: a superseded read is retried and never committed as an empty project. Proven red on base — with the old source and these tests, the project is never set (the null wins).
• projects/model.test.ts: the structure-only fallback is not content; the same node with rows, and a hydrated project whose rows were filtered away, still are.
• store/projects.test.ts: the two tests asserting the old null contract now assert the superseded one. Their intent (a late answer from the previous profile must not land; an all-profiles read must not fire) is unchanged.
• sessions-section-new-session-drag.test.tsx: its ./projects/model mock now spreads the real module, so a new export cannot silently vanish from the mock.

[thirstycrow]

Closing in favour of keeping this as a local fix on my machine.

Context for anyone arriving here later: the symptom class is already covered upstream by #104964 (merged, issue #104280) for the case where the drill-in fetch rejects, and #104301 covers the reporting side. The narrow path this PR closed is still open on top of both: fetchProjectSessions returns null for a read that was superseded (generation bump / no profile scope), which the hook commits via setProject(null) — indistinguishable from an authoritative empty project. Reproduced on the current base (reverting this PR's two source files): the superseded read resolves null instead of rejecting, and the structure-only overview node (sessionCount 43, zero lane rows) still counts as content. Diff is left intact for reference.