DR-125: scope_tag is mandatory at promotion, so the taxonomy topic grain is transitional

DR-125 — scope_tag is mandatory at promotion, so the taxonomy topic grain is transitional

Context

The S528 census could not rule on primary_domain / primary_subtopic because every route into the question dissolved into a prerequisite. S529’s lane A1 found the anchor the census had missed entirely: the pair mints the identity of topic concepts in the OKF bundle — l_records.py:926 writes topics/{domain}--{subtopic}.md, and BI-2 makes rel_path the concept’s identity, so renaming it mints a different concept. The S528 lane 1 report contains no reference to l_records, ConceptKey or topic concepts across its 506 lines.

But that anchor is the fallback grain, not the primary one. l_records.py:212-218 describes domain verbatim as “topic locator (fallback grouping) … for scope-tag-less q_a_pairs clusters”. The primary grain is scope_tag — which reference/entity-glossary.md:17 names as the q_a_pair relevance mechanism: “Relevance is computed at query time via scope_tag overlap, not a scoping FK.”

Whether the fallback is transitional or permanent turned entirely on one unanswered question: does the ratified model intend every promoted q_a_pair to carry a scope_tag? q_a_pairs.scope_tag carried no promotion rule from any ratified source, and lane A1 correctly recorded it UNDECIDABLE rather than guessing. (Corrected S531: the column is text[] NOT NULL DEFAULT '{}', not nullable as this section first recorded — so the enforcement predicate is NON-EMPTY (array_length(scope_tag,1) > 0), not SET NOT NULL. The ruling is unchanged; only the mechanism reads differently. Measured against Platform staging: 134 of 143 published pairs empty — synthetic data, recorded as shape not evidence.)

Decision

Every promoted q_a_pair must carry a scope_tag. It is a requirement of promotion, not an optional enrichment.

It follows that the (primary_domain, primary_subtopic) topic-concept grain is a transitional fallback whose requirement expires at first bundle publication, not a permanent part of the concept model.

Alternatives considered

• Leave scope_tag optional. Then the domain/subtopic pair is permanently load-bearing for bundle concept identity and needs a real enforcement mechanism in its own right — which the schema has never had (zero FK, zero CHECK, zero trigger on any taxonomy vocabulary column; reference_items.layer is the sole counter-example and carries a real validating trigger). Rejected: it entrenches a grain the ontology work is replacing.
• Defer behind a dedicated spike. Rejected: the question was already fully framed by A1’s measurement, and deferring would have left id-412 AC-10, id-414 and id-415 parked for another cycle.

Consequences

• The subject axis still mints concept identity until the fallback is gone, so the write-side contract must be fixed now rather than skipped as doomed. That work is id-419, which also carries a safety defect (validateDomain silently rewrites an out-of-taxonomy domain to validDomains[0], measured as security, at full LLM confidence) and the inert enumeration guard (AND sd.primary_domain IS NOT NULL over a NOT NULL DEFAULT 'unclassified' column, which cannot exclude a row at any corpus size).
• Promotion needs an enforcement point for the scope_tag requirement; without one this ruling is a convention, and DR-104’s lesson is that conventions unenforced by a gate regress.
• This does not by itself retire the columns. id-412 AC-10 / id-414 / id-415 stay parked until the fallback is actually removed — the filter question is settled, the columns are not yet gone.
• The measured sentinel defect (topics/compliance--unclassified.md in a staging replay) is latent, not shipped: the minted canonical-okf-showcase bundle holds four clean topic identities and no unclassified.

(S531 amendment, owner-ruled: the canonical-okf-showcase publication SATISFIES the “first bundle publication” expiry clause — the question of whether it meant a client bundle was put to the owner and ruled. The fallback grain was deleted the same session (48821e6d8: l_records enumeration + ConceptKey fields + the ?domain=&subtopic= URI form on both sides), which also removed the inert enumeration guard by construction. Consequently the parking condition on id-412 AC-10 / id-414 / id-415 — “until the fallback is actually removed” — is now MET; those tasks unpark. The write-contract fix landed as id-419; scope_tag implementation is id-422.)