DR-149: The entity_type vocabulary opens as guided-open, and a type label is added, never re-labelled

DR-149 — The entity_type vocabulary opens as guided-open, and a type label is added, never re-labelled

Context

The S554 entity-architecture review found the closed 12-value entity_type vocabulary (DB CHECK + extractor Literal + TS union — {449.1} added a fourth point, the producer lint, and a fifth column with no bound at all, entity_type_override) to be the same closed-taxonomy shape DR-141 withdrew one level up, and ruled (D11) that the direction is to open it, gated on the id-449 extraction spike because extraction had not been reviewed recently.

The spike ran (S555, {449.2}): a real-tier grid — bounded (A) vs open (B) vs guided-open (C) × 3 repetitions × the nine vendored Platform-corpus documents, scored against an owner-approved per-document reference — 108 calls, zero parse errors. C tied the bounded control on every metric (100% required recall, identical false-positive count, identical type distribution) and minted zero out-of-vocabulary labels; fully-open B minted a small semantically-reasonable tail (region, programme) but tripled source-label false positives. Every one of the 24 reference canonical groups received the same majority type in every lane, including all five contested groups.

Decision

1. The emission contract is guided-open (variant C): the twelve values are preferred labels, and the extractor may mint a new short lower-case label only when none fits, using it consistently. Type is validated for shape (short, lower-case, machine-stable), never for membership — DR-141’s rule applied at the entity level.
2. The type-stability rule (DR-147’s discipline extended to the type axis): a re-run may add labels to the corpus but may never re-label an existing entity’s type — entity_type is a component of the surrogate id and the natural key, so a flapping label is whole-corpus row churn through the type axis. A would-be re-label surfaces to the user as a suggestion (owner amendment, S555 board), exactly as DR-147 routes would-be canonical renames to curation; the pipeline never applies it silently. Probe as an AC.
3. entity_type_override is bound by the same ruling — it carries no CHECK today and is the column curators write; whatever holds for the base column holds for it.
4. The DB CHECK is replaced by a shape check, never silently dropped; the extractor Literal, the TS union and the producer lint move to the same shape contract in the same change (id-453).

Alternatives considered

• Keep the closed set as a dated deviation. Measured quality cost on this corpus: zero (C ≈ A). Rejected for DR-141 inconsistency and because the B lane already showed a real vocabulary gap (region) that a closed menu silently refuses.
• Fully open (variant B). Rejected on measurement: 3× the excluded-term false positives and a minted tail without the anchoring preference — the guided contract gets the openness without the noise.

Consequences

• id-453 implements: the C prompt contract, the five enforcement-point moves, the override binding, and the stability probe — sequenced with id-434 (the reshape’s uuid5(ns, "em:{sd_id}:{resolved}:{type}") identity and per-type resolution partitions are where a new label lands as a new partition).
• The S551 standard-vs-certification divergence is closed as a mock artefact: the real tier types the ISO/FAS/SSCM family certification unanimously ({449.2} F2); id-451’s grain family should treat certification as this family’s real-tier label.
• Resolution-partition effects of any minted label remain to be measured at the confirming /walk post-id-434 ({449.1} §6.5 metric 4) — owned by id-453.
• Honest limits carried from {449.2}: one synthetic nine-document corpus, one model (claude-opus-4-6); the stability rule, not the vocabulary bound, is what protects identity if a model change shifts typing behaviour.