commit-msg.txt

fix(subagent): populate CompletedCache in onTerminal; refactor tasks stdin TODO

## BLOCKER: CompletedCache never populated
`SubagentManager.forkSubagent` built a `CompletedCache` and exposed it via
`manager.completed`, but the `onTerminal` callback only called
`this.active.delete(id)`, `this.abortGraph.dispose(id)`, and
`void logWriter?.close()` -- `completed.add()` was never called. After
`onTerminal` fired, `manager.get(id)` returned `undefined` for every
completed subagent, making `/tasks:view` unable to locate any finished handle
from the in-memory path.

Fix: add `CompletedCache.recordHandle()` -- a convenience method that builds
a lightweight `SubagentResult` from terminal handle-state fields (all
fully populated by `run()` before `_onTerminal()` fires) and calls `add()`
internally. `onTerminal` now calls `this.completed.recordHandle(...)`.
The `message` field is intentionally omitted from the cached result because
`/tasks:view` uses `SubagentLogReader` for full conversation replay.

## Medium: subagent.ts file-size ceiling
Adding the completion-cache wiring pushed `subagent.ts` briefly above the
350 code-line ceiling. Extracting the result-building logic into
`CompletedCache.recordHandle()` brings it back to 344 code lines.

## Medium: direct process.stdin access in tasks.ts
Added a TODO comment on the `process.stdin.on('data', onKeypress)` line,
noting that direct stdin listeners bypass the compositor's routing and
should route through `compositor.enterPickerMode()` instead.

## Tests
Added 4 tests for `CompletedCache.recordHandle()` covering succeeded,
succeeded-with-stopReason, failed, and cancelled cases (17 tests total,
up from 13).