B-93 scoping — the FK census: the set is open-ended, and a third blocker is live — v0.1

Date. 2026-08-15 · Item. [B-93] (two named FKs, twice-predicting). Mode. READ-ONLY, nothing built. Engine 1adcd17 deployed; production data. The ranking, as the Operator framed it: earned by evidence — two eye-tests paid its toll, it predicted its own second instance before the fix — and sharpened by B-98's faces shipping: a conversationally-created candidate with an open finding can now be UNBLOCKED but still can't be DISCARDED. The permanent-object pair is half-closed; B-93 closes the other half. The ordered question: are two FKs the whole set, or is "handle the FKs that point at a candidate" open-ended?

The answer: 27 FKs, 4 handled, 3 observed blockers — and the set is open-ended by construction

The census: 27 foreign keys reference engagements(id) (information_schema, production). discard_candidate_engagement handles exactly 4 (memory_events, memberships, engagement_api_keys, uploaded_files) plus the row itself — a per-table list grown one incident at a time. 23 are unhandled.

Of the 23, THREE are observed blockers — the census found the third before anyone paid its toll:

1. conversation_turns — the first (E0128's shape, CR-2026-209's cleanup).
2. host_account.current_engagement_id — the second (visiting sets focus; two eye-tests paid it).
3. companion_notifications — NEW, found by this census: E0129 holds one right now. The surviving control-test candidate is undiscardable today for a reason nobody has hit yet — the endpoint would 500 on companion_notifications_engagement_id_fkey before ever reaching the turns it would also 500 on. The suppression property's third prediction, caught by enumeration instead of by toll.

The remaining 20 hold zero candidate rows today — and "zero today" is not "unreachable." Candidates run conversation, rooms, and job machinery (notifications just proved a "surely active-only" table reachable); any future feature that writes engagement-scoped rows during candidacy silently adds a member. The set is open-ended by construction. Two was never the number; the number changes with the schema.

What this decides about the CR's shape

The per-table DELETE list is the wrong shape — it is how the endpoint got here (4 entries, each added after a block). The right shape is the arc's own pattern, fourth application (after CR-204's AST test, CR-205's source-scrape, CR-213's kinds fence): the contract enforced against the source — here, the schema.

• The discard declares a policy for EVERY referencing FK: delete (rows about a thing being unmade — turns, notifications, uploads, events, memberships, api keys), null (stale pointers — current_engagement_id), or active_only (a declared assertion that this table cannot reference a candidate — render pipeline, manifestation views, contributors-at-commit, personal_engagement_id).
• The fence test reads the schema (information_schema / SQLAlchemy metadata) and fails when a referencing FK exists with no declared policy — a new FK added anywhere in the system without a discard decision is a failing test, not a 500 on someone's cleanup months later. active_only declarations are cheap and honest; a wrong one converts to a one-line policy change the day its table proves reachable.
• Policies for the observed three: turns → delete; focus → null; notifications → delete. The other 20 get their declarations decided at drafting — most are plainly active_only; the drafting cost is reading, not designing.

Fix-shape delta from B-93's entry: the entry's "sweep information_schema so the fix enumerates rather than chases" is confirmed and PROMOTED from a drafting step to the mechanism itself — the sweep is not preparation for the fix; the sweep-as-standing-test IS the fix's load-bearing half, exactly as the kinds fence is CR-213's.

Also established

• E0129 is the live specimen: 2 turns + 1 notification (+ focus whenever visited) — three blockers on one disposable candidate. Its discard is the natural eye-test.
• The B-98 pairing update for the entry: the pair is half-closed as the Operator states — unblockable now (CR-212's panel), undiscardable still. B-93 alone completes "no permanent objects."

What awaits ruling

The B-93 CR draft on the policy-plus-fence shape (engine only; the endpoint, the policy table, the schema-reading fence test, the three observed policies, twenty declarations; E0129 as the eye-test specimen, which also finally cleans up the control-test residue).

DUNIN7 — Done In Seven LLC — Miami, Florida — B-93 FK census — v0.1 — 2026-08-15 Two was never the number. Twenty-seven references, one policy each, and a fence that asks the schema — the census caught the third toll before anyone paid it.