Add FetchedMemoriesChip feature flag, AIConversation::fetched_memories()
accessor, the FetchedMemoriesChipView popup, footer toolbar wiring, and
telemetry. Patches Cargo to the local warp-proto-apis for the new field.

Co-Authored-By: Oz <oz-agent@warp.dev>

- New AIAgentCitation::AgentMemory { memory_store_id, memory_id } variant
- TryFrom<api::Citation> splits compound 'store_id:memory_id' document_id
- AgentModeCitation::AgentMemory added to telemetry events
- for_telemetry impl maps AgentMemory through to the telemetry type
- render_citation: Cognition icon + 'Memory' label for AgentMemory chips
- OpenCitation handler opens oz_root_url/memory/{store}/memories/{id}

Co-Authored-By: Oz <oz-agent@warp.dev>

Newer local proto branch adds fields/variants not yet handled in client:

Struct literals — add ..Default::default() to:
- api::request::Settings (custom_model_routers, supports_background_computer_use)
- api::message::ModelUsed (prompt_cache_expires_at)
- stream_finished::ConversationUsageMetadata x2 (context_window_segments, total_input_tokens)

Exhaustive matches — add WaitForEvents arms to:
- write_tool_call_args (conversation_yaml.rs) — no-op, no args to serialize
- write_tool_call_result_content (conversation_yaml.rs) — emit 'status: completed'
- convert_tool_call_result_to_input (convert_conversation.rs) — return None
- create_cancelled_result_for_tool_call (convert_conversation.rs) — return None
- ToolExt::name (task/helper.rs) — 'wait_for_events'

Co-Authored-By: Oz <oz-agent@warp.dev>

- Add content: String to AIAgentCitation::AgentMemory so chips show a
  truncated preview of the memory text (like rules show their title)
- Hash/Eq implemented manually to key on IDs only, not content
- render_citation shows truncated content or falls back to 'Memory'
- At render_references_footer call site, synthesize AgentMemory
  citations directly from conv.fetched_memories() -- no LLM XML step,
  no timing issue, works for live/restored/cloud conversations

Co-Authored-By: Oz <oz-agent@warp.dev>

…d_output

render_citation_chips gates rendering on citation_state_handles.get(citation).
Memory citations are synthesized at render time from fetched_memories, so they
never go through the output.citations path and never got handles registered.

Fix: in handle_updated_output, also register AgentMemory handles from
conversation.fetched_memories() alongside the existing output.citations loop.
The entry().or_default() pattern is idempotent so repeated calls are safe.

Co-Authored-By: Oz <oz-agent@warp.dev>

Memory retrieval only happens for the first user message, so the
References section should only show AgentMemory chips on the first
exchange's output block, not on every subsequent turn.

Use conv.first_exchange().id == exchange_id check at both the render
site (output.rs) and the handle registration site (block.rs).

Co-Authored-By: Oz <oz-agent@warp.dev>

Keep fetched_memories proto field and AIConversation::fetched_memories()
since they power the References section citations. Remove everything
else: FetchedMemoriesView, FetchedMemoriesChip feature flag, the
AgentToolbarItemKind::FetchedMemories variant, and its toolbar wiring.

Co-Authored-By: Oz <oz-agent@warp.dev>

Co-Authored-By: Oz <oz-agent@warp.dev>

- Apply rustfmt formatting to citation.rs and output.rs
- Remove duplicate WaitForEvents match arms in conversation_yaml.rs and
  convert_conversation.rs that caused unreachable pattern clippy errors

Co-Authored-By: Oz <oz-agent@warp.dev>

AgentMemory citations are derived from fetched_memories at render time,
not from the Citation proto, so TryFrom<api::Citation> doesn't need to
handle this case.

Co-Authored-By: Oz <oz-agent@warp.dev>