Stop the tree flickering on zombie intermediate parents; skip dead sessions in parent auto-detect

[curusarn]

Problem

When an intermediate parent in the session hierarchy is a zombie (killed session that live sessions still reference as parent, e.g. crab-t-10026-worker referenced by crab-u7449-conflict), the TUI session list flickered between two shapes:

Shape A (correct) — dead parent as a dimmed virtual root:

>        t-10026-worker                       -   -
  local  ├── crabctl-zombie-parent  running
  local  └── u7449-conflict         waiting
  local  orchestrator               waiting

Shape B (glitched, ~1.5s after every remote poll) — the zombie re-nests under its own recorded parent, children pushed a level deeper, roots reorder:

  local  orchestrator                             waiting
>        ├── t-10026-worker                       -   -
  local  │   ├── crabctl-zombie-parent  running
  local  │   └── u7449-conflict         waiting

Root causes

1. Flicker: the remoteSessionsMsg handler (and the ctrl+h fold path) re-feed BuildTree its own previous output, which contains the synthetic Virtual row for the dead parent. Inside BuildTree that stale row registered as an active session, so it stopped being a virtual root, picked up its own DB parent, and re-nested its subtree. The next local refresh dropped it again — flicker on alternating refreshes. Fix: BuildTree drops Virtual rows from its input (they are derived placeholders, always re-created).

2. Zombie edges keep being created: DetectParent trusted CRABCTL_NAME (which outlives the tmux session it came from) with no liveness check, so spawns recorded parents that had been dead for an hour. Fix: all auto-detect call sites (crabctl new, TUI /new, dir-picker, orchestrator spawn) reroute a dead detected parent to its nearest live ancestor via the recorded parent chain (top level if the whole chain is dead, cycle-safe). Explicit -p is still trusted as-is.

Tests

• TestBuildTreeZombieIntermediateParent — A live ← B dead ← C live: B renders as virtual root, C nested, no rows lost
• TestBuildTreeRebuildFromOutputStable — the regression: rebuilding from previous output is shape-stable (failed before the fix with exactly the depth shifts seen live)
• TestBuildTreeWholeChainDead — only the direct dead parent materializes
• TestNearestLiveAncestor — live candidate, dead→live walk, all-dead chain, cycle safety, unknown candidate

go test ./... and go vet ./... green.

Verification against the live tmux server

• Before: flicker to shape B captured within 40s of watching.
• After: 10 frames over 60s, shape stable; virtual root never re-nests.
• Facet 2: CRABCTL_NAME=crab-t-10026-worker crabctl new zzz-test-child (stale env pointing at the dead session) resolved Parent: crab-orchestrator — the nearest live ancestor.

🤖 Generated with Claude Code