[Bug] @ mention picker is flooded with subagent worker transcripts (3-line fix proposed)

[Minokun]

What happened

The @ reference picker in the web input lists every persisted session — including
delegation worker sessions (origin: "subagent", non-null parentSession). After one
heavy orchestration run (~57 parallel review subagents in my case), the picker became
unusable: dozens of duplicate entries titled by the worker prompt
("你是计划审查员" × N, one per review arm per sample).

Why it happens

SessionReferenceResolver.listCandidates()
(packages/context/session-reference/src/index.ts) excludes only self:

.filter(record => record.header.id !== agent.id)

parentSession is already indexed (session-query-sqlite stores it, filters.ts
can filter on it) — the candidate list just never uses it. Worker transcripts are
not conversations a human references; their content is reachable through the parent.

Proposed fix (3 lines + test)

     const records = (await settleWithCancellation(this.ctx.sessionQuery.listSessions(signal), signal))
+      // Subagent child sessions (delegation workers) are transcripts, not
+      // conversations a human references; exclude them from mention candidates.
+      .filter(record => record.header.id !== agent.id && record.header.parentSession == null)
       .map((record, index) => ({ record, index }))

With a regression test creating sessions with parentSession set and asserting they
never appear in listCandidates() output. Package suite passes 17/17 with the patch.
I can't submit a PR at the moment (external PRs paused per CONTRIBUTING) — happy to
post the full diff + test here if useful, or re-submit when PRs reopen.

Workaround meanwhile

Archiving worker records out of ~/.dsh/sessions cleans the picker
(records are plain directories; verify origin: "subagent" in the first JSONL line
before moving).