fix(reflection): count canonical assistant steps

[just-cameron]

Overlord (agent-c2adbf5c-8419-4211-8cd8-3740db164974)

Problem

A single backend assistant message can become several transcript rows when the CLI splits long output for display. Reflection cadence counted each row as a completed step. One long response could therefore reach the configured step threshold early.

Change

• Count each non-empty backend assistant message ID once per completed turn.
• Keep one-row-per-step behavior when a transcript row has no backend message ID.
• Migrate older reflection state from transcript history.
• Clear legacy cursors that no longer exist, so migration cannot skip unreflected content.
• Advance reflected cursors past all rows that share the same canonical message ID.
• Preserve every transcript fragment in reflection payloads.

Normal completed-turn accounting remains O(delta). Full transcript scans occur only for migration, external-source filtering, and payload/finalization paths that already need transcript history.

Validation

• Regression test uses the accumulator streaming path to create two display rows with one backend message ID.
• The regression suite fails against the parent implementation and passes with this change.
• Focused reflection suites: 43 passed, 0 failed.
• bun run check: 12/12 checks passed.
• Independent adversarial review found one missing-cursor migration case; the branch includes the repair and a regression test.

Risk

This changes local reflection state schema from v3 to v4. Existing v2 and v3 state rebuilds counters from transcript history. Missing canonical IDs retain the prior row-count behavior.

👾 Generated with Letta Code