test - #4

new wails bindings format working

Complete three subagent-tree v1 follow-ups on the shared disk-backed data
layer (GetSubagentsCommand -> SubagentFileInfo -> correlateSubagents ->
subagentsByIdAtom), so the disk store is the single source of subagent truth.

Shared foundation:
- Add a neutral "done" SubagentState (finished, outcome unknown) with a muted
  color in the tree, Runs dot, and card peek.
- Extract useSubagentTracking() from the tree so the grid and Runs load the
  store the same way instead of copying the effect.

(lastRecordTerminal) into SubagentFileInfo.Done; correlateSubagents resolves a
terminated orphan (no parent Task tool_result) to "done" instead of a perpetual
"working". Success/failure stays reserved for the parent's real judgment.

rows now read subagentsByIdAtom; the AgentSubagentDelta emission, the
agentstatusstore reducer/TTL/idle-clear, getSubagentsAtom, and
baseds.AgentSubagentDelta are deleted. (getSubagentsAtom was NOT dead -- it was
live in the rail + Runs -- so it was migrated first, then removed.)

listing each child's type + state dot; clicking opens the focused view.

Plan under-enumerated the deletion: also removed the subagent-delta emitter from
wsh agentstatus (wshcmd-agentstatus.go) and updated wshcmd-agenthook_test.go +
sessionviewmodel.test.ts, all forced by the go build / tsc gates.

Closes wavetermdev#2 (Codex has no per-subagent files) and wavetermdev#4 (0 nested subagents dirs
across 619 real child files) as no-go, recorded in deferred.md.

Spec/plan: docs/superpowers/{specs,plans}/2026-07-10-subagent-tree-followups*.

First tranche of Theme 4 (maintainability & test gaps), executed on current
main after Theme 1 and Theme 3 landed.

Test-gaps (characterization; each mutation-verified):
- runactions.test.ts — confirmCancelRun live-worker branch/copy, cancelRun +
  stopRunWorker in-flight tracking and finally-cleanup, tab: strip.
- agentcwdresolve.test.ts — resolveCwd block->tail->head precedence; head read
  fires only on a tail miss; null at the error boundary.
- pkg/jarvis/watcher_test.go + onexit_test.go — extracted pure predicates
  askAutoAnswerable (single single-select pre-filter) and optionIndexInRange
  (delivery bounds guard) from handleAsk (behavior-preserving) and tested them;
  outcomeSummary last-event-vs-task + 160-char truncation. wavetermdev#2 was unblocked by
  Theme 3 A1 (dac43d1).

Refactor wavetermdev#6 (behavior-preserving):
- findSessionTermBlock pure selector added to sessionviewmodel.ts; all 4
  duplicate "first term block with cmd:cwd" sites in sessionsidebarmodel.ts
  routed through it via a resolveTabBlocks id-resolver.

Deferred: wavetermdev#4 (runbody.tsx split) and wavetermdev#5 (agentsviewmodel.ts grid extract) --
pending the active Theme 2 work that edits those files.

Verified: vitest run (1138 pass), large-stack tsc (clean), go test
./pkg/jarvis/ (ok), go vet clean.

…vetermdev#4/wavetermdev#5

Work through docs/open-issues.md; wavetermdev#5 deferred (documented blocker).

wavetermdev#1 gitattributes merge=union for the three generated wshrpc client files
   (wshclient.go, wshclientapi.ts, gotypes.d.ts); tsgen guards backstop a bad union.

wavetermdev#2 split two agent-view god-files (move-only, call sites unchanged):
   - runbody.tsx 881->561: card family -> runcards.tsx, plan editor -> planpreview.tsx
   - agentsviewmodel.ts 1030->895: grid geometry -> cardgridlayout.ts (+ tests),
     re-exported so importers are untouched.

wavetermdev#3 (3a) transcript-stream per-card unmount watcher leak:
   - server cancelRequest now cancels the request context (not just the bool), so a
     streaming handler's goroutine + fsnotify watcher unwind while the link is up.
   - client emits the wire cancel synchronously from an overridden return() in
     sendRpcCommand -- gen.return() hangs and never runs finally when the generator is
     parked at the never-settling await (a quiet stream), so a finally-based cancel
     would be missed. Guarded by go -race + client vitest.
   - 3b (stateful projection) intentionally not built (measure-first).

wavetermdev#4 channel-attachment temp-file cleanup: WriteTempFileCommand now uses a distinct
   waveterm-attach- prefix (the bare waveterm- sweep could delete /tmp/waveterm-<uid>
   socket dirs on Linux/macOS); SweepTempAttachments reaps dirs >24h, wired into
   wavesrv startup + a 4h loop.