claude-fork-trace-identity

# Give each forked Claude Code session its own trace

The backend's local trace uploader (`backend/app/agents/traces/local.py`, run
as `just backend jobs upload_local_traces`) publishes a developer's Claude Code
sessions from `~/.claude/projects/<project>/<file>.jsonl` to the trace bucket,
one capture per session, replaced in place as the session grows. It names a
Claude Code session by the `sessionId` in its transcript.

That loses work. When Claude Code forks or resumes a conversation, it writes a
new transcript file, named by the new session's id, whose leading records are
copies of the parent's, `sessionId` included; only the records written after
the fork carry the new id. Every fork of a conversation therefore uploads under
the parent's session and trace id, and each upload replaces the last.

Make each transcript file publish as its own session:

- A Claude Code session's id is its transcript file's name without `.jsonl`.
  Its trace id is derived from that id the same way as now, so the session a
  conversation began in, whose file name is its `sessionId`, keeps the trace id
  it has today.
- When the `sessionId` a file's records start with differs from the file's
  name, the capture records that inherited id as its `parent_thread_id`;
  otherwise `parent_thread_id` stays null.
- Everything in a fork's capture names the fork as its session: the
  `native_session_id` in its manifest and in the archive's discovery metadata,
  and its normalized records, including those copied from the parent.
- Claude Code captures from cloud workers (the worker harness's Claude Code
  trace export) must come out exactly as they do now. Codex sessions are
  unaffected.

## Environment

PostgreSQL on localhost:5432 and the local database endpoint the app's
driver uses are running when you start, and the dev env files point at the
dev database.