Bring the Claude transcript projection to server parity #775

Part of #679. Second of five steps; the rules are the Claude section of docs/superpowers/specs/2026-09-04-ai2265-transcript-normalization-leaf-design.md.

Deliverable

ClaudeTranscriptEvents in Capacitor.Models.Transcripts produces everything the server's ClaudeCodeNormalizer produces today, so the server can call it instead:

• usage from message.usage on the record's first emitted event, governed by the previous-usage memory with the transitions the spec states (clear on any other record, untouched without usage, compare-then-store with usage even when nothing is emitted);
• tool-result text per the server rule: a string as is, an array's text blocks joined with a blank line or its raw JSON when it has none, any other content as raw JSON, unset when absent; extensions.claude_code with tool_use_result, output_raw, is_error;
• image blocks as TranscriptAttachments with the scoped attachment id; attachment records of type queued_command as user messages;
• is_meta on user messages, is_sidechain on every event of a sidechain record, origin_kind on task notifications.

Done when

• The leaf's synthetic Claude fixtures and golden files cover every rule above, including multi-block assistant records, multi-result user records, an invalid uuid, and the usage-memory sequences.
• Identifier vectors include the attachment id.
• The Chat tab's output is unchanged (Core's ClaudeChatRulesTests stay green).