📖About

Goal (Ninad): make the bypass methodology (Orion + Albus) as close to planner as possible. Today a bypass only writes a hollow start/end bookend to agent_activities with NO initiative_id and no content, so: (a) closed bypass initiatives never appear in the Activity Tracker (it excludes initiatives with 0 activities, counted from initiative_metrics which is a stale backfill), and (b) the work is invisible. Fix in shared/bypass.py (used by BOTH Orion and Albus bypasses): at close, reconstruct the REAL work as planner-equivalent agent_activities rows tied to the initiative_id — a code row (actual files changed since the bypass opened, from git), a decision row per recorded initiative decision, and a review row (twin-reviewer verdict) — then upsert initiative_metrics live (activities = real count, qualifying=1) so the tracker shows genuinely-worked bypass initiatives exactly like planner runs. No dummy data — every row derives from a real artifact. Forward-fix.

⚖️Decisions

• Caller audit waived by Orion: additive close-time recording, no contract impact — CAROL-INI-1840 adds two helpers + one best-effort call at the tail of bypass_end (after the close work). bypass_end signature, return shape, and all side-effects for callers (orion.bypass_cli, elrond.bypass_reviewer, elrond.initiative_reviewer, apps.initiatives.app) are unchanged; the new code only writes agent_activities + initiative_metrics, wrapped in try/except so it can never raise. No caller reads the new rows for control flow. Safe. (orion)

✅Success criteria

• A bypass records real per-block activities (code/decision/review) tied to its initiative_id (must_have)
• bypass_end upserts live initiative_metrics so the Activity Tracker shows the closed bypass initiative (must_have)