feat(chat): show which stored memories informed an answer

[sulthannauval]

Problem

The agent recalls stored memory into every prompt without the user asking, and
the console said nothing about it. An answer leaning on a remembered fact was
indistinguishable from one the model simply knew — and there was no way to find
the entry responsible.

Change

• memory_recalled on the chat stream sets recalledMemories on the assistant
  turn.
• Rendered as a chip row beside Sources — both answer "what informed this?".
• Auto-saved turns are counted, not named. Their keys are
  user_msg_<uuid>; the first version rendered five of them and filled the row
  with hex while identifying nothing. Named memories are still named.

Depends on

RantAIClaw #417 (AgentEvent::MemoryRecalled → memory_recalled SSE frame).
Without it the chip row simply never appears.

Validation

npx tsc --noEmit clean, npm test — 156 passed (20 files), including new
coverage for the summarisation and for telling a generated key from a chosen
one that merely contains underscores (plan_2026_08_06).

Driven in a real browser against a live gateway, with a stub provider so the
turn completes:

Atlas  LLAMA3
Deployments happen Friday afternoons only.
RECALLED:  ⟐ 5 from this conversation

The stub matters: every earlier attempt failed at the provider, and a failed
turn drops out of the transcript — so "no chip" looked like a render bug when
it was just a turn that never finished. The plumbing was confirmed separately
by reading the frame off /api/chat before trusting the UI.

Risk / rollback

Low: one new optional field, one conditional chip row, one pure helper. Revert
the commit. Against an older gateway the event never arrives and nothing
renders.