TECH — {427.3} Grain registry, residual grain, and the retirement of the closed type vocabulary

TECH — {427.3} OKF producer inversion

Task: id-427. Artefact: {427.3} TECH. Date: 09/08/2026 (S546). Reads: RESEARCH.md {427.1}, PRODUCT.md {427.2}. Paths: producer scripts/cocoindex_pipeline/producer/; Source adapters scripts/cocoindex_pipeline/sources/. All line numbers are main d1e6e14ad.

---

1. Architecture — the one structural move

Today the producer has seven enumeration branches and three separate dispatchers keyed on concept_type (read_concept l_records.py:993-1017, _source_documents_for_key :1056-1085, sample_rows :1185-1210), plus a directory literal embedded in each branch, plus a write-time redirect for the one case where two branches collided. Adding a grain means editing four places and, today, also editing four type registries. That structure is why nobody added a catch-all.

The move: a grain is a declared object, and everything about it lives in one place.

@dataclass(frozen=True)
class GrainSpec:
    name: str                  # dispatch key, e.g. "topic_scope_tag"
    directory: str             # bundle directory this grain owns, e.g. "topics"
    type_label: str            # the `type:` value concepts of this grain carry
    list: Callable[[Pool], Awaitable[GrainEnumeration]]
    read: Callable[[Pool, ConceptKey], Awaitable[ConceptRaw]]
    sample: Callable[[Pool, ConceptKey, int], Awaitable[list[Mapping[str, Any]]]]
    drafts_via: Literal["pass1", "template"] = "pass1"

@dataclass(frozen=True)
class GrainEnumeration:
    keys: tuple[ConceptKey, ...]
    covers: Coverage           # the unit ids this grain's concepts reach

@dataclass(frozen=True)
class Coverage:
    source_document_ids: frozenset[str] = frozenset()
    q_a_pair_ids: frozenset[str] = frozenset()

ConceptKey gains grain: str (the dispatch key) alongside concept_type (the emitted label). list_concepts() iterates the registry, unions every grain’s Coverage, and then runs the residual grain last, handed that union. read_concept/sample_rows become self._grains[key.grain].read(...) / .sample(...). _source_documents_for_key stops sniffing concept_type and takes an explicit patterns argument from its grain.

Consequences, all deliberate:

• Adding a grain is one registry entry. No dispatcher edit. This is the property whose absence produced the inversion.
• concept_type is no longer read by any control flow — only written into frontmatter (enrich.py:939) and into the type filter the viewer derives from the bundle (bundle-graph.ts:515). That is what “a label, not a gate” means mechanically.
• The directory is the grain’s, not the type’s (RESEARCH M5/C2). Relabelling a concept cannot move its file — PI-5, and BI-2’s memo key stays put.
• repo_docs.RepoDocsSource uses the same registry shape for its two pillars, so the two Sources stop being parallel implementations of the same idea (id-362 F1).

---

2. Decisions

2.1 The residual grain — coverage, not a catch-all list

Coverage cannot be computed by a standalone SQL predicate. The product / certification / case_study grains select documents by %<entity canonical_name>% ILIKE patterns that are data-dependent (l_records.py:455-466), so “the set the preferred grains reach” is not expressible as a static predicate. Therefore:

Every grain declares what it covers, and the residual grain is the complement.

Each preferred grain’s list returns, alongside its keys, one set-based query’s worth of covered unit ids (six extra set-based queries per run — no per-concept round-trips, MD-5’s existing discipline). The residual grain then issues two anti-joins:

-- residual documents
SELECT id, filename, logical_path, created_at, updated_at
FROM source_documents
WHERE publication_status = 'published' AND id <> ALL($1::uuid[]) ORDER BY id;

-- residual pairs (attribution inputs only)
SELECT id, source_document_id, source_form_instance_id
FROM q_a_pairs
WHERE publication_status = 'published' AND id <> ALL($1::uuid[]) ORDER BY id;

The invariant is coverage (≥1), not partition (=1). Measured (RESEARCH C4), a published pair with a scope_tag and a pattern-matched parent already lands in two concepts today, and that overlap is good — the same knowledge reachable two ways. DR-141’s Decision says “exactly one”; the implementable and product-load-bearing invariant is “at least one”. Action: recommend a DR-141 rider recording coverage-not-partition. This spec does not silently reinterpret the register.

2.2 The attribution cascade — three homes, one rule

A residual document is its own concept. A residual pair is attributed by a cascade, because q_a_pairs.source_document_id is nullable and is NULL for every derived_from_form_response pair (RESEARCH M6):

residual_home(unit) =
    unit is a source_document        → documents/
    unit.source_document_id          → documents/           (joins that document's concept)
    unit.source_form_instance_id     → questionnaire-responses/
    otherwise                        → unattributed-answers/ (one bundle-wide concept)

Home	Key	rel_path	type	Populated by
documents/	source_document_id	documents/<slug>-<sd_uuid[:8]>.md	document	hole 2, and hole-1 pairs that have a parent document
questionnaire-responses/	form_instance_id	questionnaire-responses/<slug>-<fi_uuid[:8]>.md	questionnaire_response	published derived_from_form_response pairs from non-won forms
unattributed-answers/	— (singleton)	unattributed-answers/published-answers.md	answer_set	published pairs with neither lineage; omitted entirely when empty

Why document is the type label. OKF §4.1 asks for descriptive and self-explanatory. residual/unrouted describe our pipeline, not the knowledge. source_document borrows a table name and invites exactly the content_type/concept-type conflation DR-050 forbids. document is what the thing is, and a generic OKF consumer reads it without a glossary.

Why the -<uuid[:8]> suffix is unconditional. write_bundle refuses a physical write-path collision before any write in the run (bundle_writer.py:1310-1320), a residual concept has no curated name to disambiguate with, and a conditional suffix would make one concept’s identity depend on the existence of an unrelated row (deleting the collider would rename the survivor — a spurious moved). An 8-hex prefix is not a record pointer: contains_record_pointer matches only the full 8-4-4-4-12 form (resource_uri.py:65-68), so BI-10 is untouched.

Title guard. The residual title derives from filename. If that filename embeds a full uuid (pipeline sidecars are minted from sd:<rel_path>), build_concept_frontmatter would raise on the BI-10 check (frontmatter.py:258-264). The renderer runs contains_record_pointer on the candidate title first and falls back to a neutral "Undistilled source document" when it hits.

2.3 What Pass-1 receives, and when it does not run

read_concept for a residual document returns the document row, its published q_a_pairs (possibly none), its record_lifecycle, entity_mentions and entity_relationships. It does not return body text: source_documents.extracted_text is permanently NULL on the pipeline path and composing content_chunks bodies is ruled out of scope (l_records.py:383-388).

So the grain declares its drafter:

• drafts_via="pass1" when the cluster has ≥1 published q_a_pair — real content, the ordinary agent loop, no special case.
• drafts_via="template" when it has none — render_undistilled_draft(key, raw) returns a ConceptDraft directly, bypassing the agent loop entirely:

# {title}

No published answer has been distilled from this document. It is held in the corpus and
is reachable here so that its absence from the answer set is visible rather than silent.

## What is known
* Document: {filename} ({content_type})
* Held since: {created_at}; last updated: {updated_at}
* Extraction: {extraction_method}
* Entities mentioned: {canonical_names}      <!-- omitted when none -->

## What is not known
No question–answer pair published from this document exists. Any question this document
might answer is unanswered by this bundle. Escalate to a subject-matter expert.

Frontmatter: type: document; deterministic description; confidence: no-content — the A19 value that has been unreachable since it was ratified (RESEARCH M8), and the honest one here; tags: []; no top-level resource: (id-426 F2-B); sources[] carrying canonical://source_documents/<uuid>; generated.by = the producer actor string per §7 — the producer is the actor, there is no human: claim to make.

Rejected — drafting it through Pass-1 anyway. The model would receive a filename and metadata and be asked for a title, description and body. It would produce plausible prose about a document it has not read. That is the exact failure the negative answer exists to prevent (PRODUCT PI-4). It also spends an Anthropic call per undistilled document.

2.4 Directory scheme, and the id-429 interface

Rule: the directory is declared by the grain. There is no type→directory function.

The existing five directories are already grain constants (RESEARCH M5) and are therefore unchanged: topics/, products/, company/, certifications/, case-studies/. The won-bid grain owns case-studies/won-bid/ (§2.5). Residual grains own the three directories in §2.2. A feeder-declared grain names its own directory (§2.7).

Rejected — directory = slug(type). It re-instates the inversion at the filesystem layer (the label would decide the location), it is not a total function over an open vocabulary (the feeder’s own docstring, l_records.py:805-812, already ruled that an arbitrary client-chosen type name has no principled English-plural rule), and it would make a relabel a file move — churning BI-2 identity, the cocoindex memo key, every BI-9 cross-link and the client’s git history.

Against id-429’s interface assumptions:

id-429	Status here
IA-1 every concept has a directory	Satisfied — every grain declares one; nothing is written at the bundle root.
IA-2 the residual grain gets a named directory	Satisfied — three of them (§2.2).
IA-3 index/log refused as concept slugs	Specified — §2.6. Net-new, created by this task.
IA-4 deterministic, stable, filesystem-safe directory names; many-to-one fine	Satisfied — they are module constants, not derived values.
IA-5 every intermediate directory is real	Satisfied — case-studies/won-bid/ is the only nesting, and it is a real directory today.

id-429 Q2 is answered, not carried. The residual directory names are readable words, so id-429 D6’s sentence-cased basenames give “Documents”, “Questionnaire responses”, “Unattributed answers” with no additional label. The opaque -<uuid[:8]> suffix appears only in file slugs, never in a directory name, so no heading is ever derived from it.

2.5 The won-bid redirect collapses

_won_bid_case_study_redirect / bundle_write_path / bundle_write_path_for_key (bundle_writer.py:224-284) exist solely because two case_study grains were forced to share one directory (RESEARCH M13). Once the won-bid grain declares directory="case-studies/won-bid", identity rel_path is the physical path and all three functions delete; flow_def.py’s BI-28 map and the embed lookup key on rel_path directly.

The physical path does not change (case-studies/won-bid/<slug>.md already). What changes is the concept’s identity, and therefore its memo key and its BI-9 citation key. No file moves; RunSummary.moved stays empty. The affected concepts re-draft once — folded into the single wave-wide version= bump (§5).

2.6 Reserved-name slug guard (id-429 IA-3 — net-new)

OKF §3.1 reserves index.md and log.md at any directory level; they MUST NOT be concept documents. Under a closed vocabulary this was unreachable. Opening it makes a scope_tag, entity name or filename that slugifies to index or log reachable for the first time, and such a concept would silently overwrite that directory’s index and then be reconciled away on the next run.

Rule — deterministic rename, not refusal. A new mint_concept_slug(value) -> str wraps _slugify (l_records.py:676) and appends -concept when the result case-folds to a reserved stem:

_RESERVED_CONCEPT_STEMS = frozenset({"index", "log"})
mint_concept_slug("Index")  -> "index-concept"
mint_concept_slug("log")    -> "log-concept"
mint_concept_slug("iso-9001") -> "iso-9001"        # identity on everything else

Why rename rather than refuse: a client document legitimately called “Index” is a data fact, not a configuration error; aborting a whole producer run over it violates DR-047’s narrowly-scoped degrade posture. Why not a _ prefix: leading-underscore files are skipped or hidden by enough tooling that it trades one silent invisibility for another. Residual collision: a directory holding both index and a genuine index-concept trips write_bundle’s existing pre-write collision guard (:1310-1320) and fails loud before any write — the correct outcome for a genuine ambiguity, and it costs no new machinery.

Directory names are guarded differently. index/log as a feeder-declared directory is a configuration error, not a data fact, so _validate_concept_feeder_schema (bundle_writer.py:918) rejects it at read time with the existing fail-loud posture.

2.7 The feeder branch under the open model

_list_feeder_concepts (l_records.py:755-821) had two halves. The type-widening half (concept-feeder.json declares a type, _permit_overlay_concept_types lets it past the ConceptKey guard) dies with the guard. The grain half — enumerate concepts over entity_mentions of entity type X — survives, and it is now the only thing the feeder is: a client-declared preferred-routing grain, which is precisely DR-141’s model.

Config gains an explicit directory key, because type and directory decouple:

{ "framework": { "grain": "entity_mention", "entity_type": "framework",
                 "directory": "frameworks" } }

directory defaults to slug(type) when omitted, for compatibility. No shipped bundle carries a concept-feeder.json (checked: none in canonical-okf-showcase/), so there is no config to migrate — this is an artefact fact, not a “0 rows” argument about correctness.

2.8 type shape validation — the exact rule

check_type_membership (validator.py:406-420) becomes check_type_shape. The name must change: an identifier that says “membership” is the inversion.

_TYPE_SHAPE_RE = re.compile(r"^[a-z][a-z0-9]*(?:_[a-z0-9]+)*$")
_TYPE_MIN_LEN, _TYPE_MAX_LEN, _TYPE_MAX_WORDS = 3, 40, 4
_RESERVED_TYPES = frozenset({"q_a_pair"})   # BI-3, unconditional

A type is valid iff: it is a str; it matches _TYPE_SHAPE_RE (lowercase ASCII snake_case, leading letter); its length is 3–40; it has at most 4 underscore-separated words; and it is not in _RESERVED_TYPES.

Why this shape. (i) iri_projection.slug() is identity on it (:132-148), so no term is silently rewritten downstream; (ii) all ten types the platform emits today — topic, product, company, certification, case_study, schema, tool, api, navigation, playbook — already match, so the rule ratifies existing output rather than inventing a migration; (iii) api is 3 characters, which sets the floor and excludes the uninformative 1–2 character token; (iv) the word cap keeps a type a label rather than a sentence, which is what “self-explanatory” buys.

Deliberately not checked: singular-vs-plural (not objectively decidable, and it would reject nothing a human would call wrong); collision with a content_type value (DR-050 forbids merging the vocabularies, not string coincidence — adding that check would be the overreach it warns against).

The effective_ontology parameter is dropped from this check: there is no set to compose against. check_concept/validate_concept still take effective_ontology for lint_entity_relation_mentions, which is unchanged.

2.9 The four registries + the facet vocabulary

Registry	Disposition	Requirement it served, and why that ends
ALLOWED_CONCEPT_TYPES validator.py:116-118	DELETED	BI-4’s write gate. Withdrawn by DR-141 / DR-019 amendment.
CONCEPT_TYPES l_records.py:115-117 + guard :273-284	DELETED	Source-side defence-in-depth on the same set. Guard replaced by the §2.8 shape check + the BI-3 q_a_pair refusal, which stays unconditional.
_CLASS_CONCEPT_TYPES validator.py:242-248	DELETED	DR-079’s per-class type sets. Owner ruling (iii), S546 — uniform across bundle classes. Downstream: EffectiveOntology.base_for_class and repo_docs.SYSTEM_BASELINE_CONCEPT_TYPES delete with it.
CONCEPT_TYPE_VALUES lib/ontology/concept-schema.ts:70	DELETED	Its own docstring sources it only to “documentary/UI purposes (e.g. a future type-legend)”. Measured (RESEARCH M4): no runtime consumer, and the legend duty is already discharged by concept-type-tokens.ts KNOWN_TYPES. A future use is not a current source.
KNOWN_TYPES lib/okf/concept-type-tokens.ts:38-48	KEPT — the one surviving legend	Colour tokens with an explicit default fallback for unmapped types. Extended with document, questionnaire_response, answer_set, reference. Additive; never a gate.
RECOGNISED_FACET_TAGS validator.py:147-149	DELETED	Its own two stated requirements (:120-133): (a) name first-class facets for downstream consumers — discharged by KNOWN_TYPES; (b) stop the type set and facet vocabulary diverging — subject stops existing. Facet tags themselves survive as ordinary free-form tags: entries; they never needed a registry (check_concept has never rejected an unregistered tag).
FACET_TAG_ALIASES + canonical_facet_tag + normalise_facet_tags :157-175	DELETED	The methodology → playbook fold existed because the type set would not admit a sixth type — DR-141’s headline evidence. No production caller (RESEARCH M3).
_permitted_overlay_concept_types + _permit_overlay_concept_types l_records.py:152-168	DELETED	Existed only to lift the ConceptKey gate for feeder types. Gate gone, widener gone.

The three bends, resolved as types (AC 3):

• reference — the only one with a live emitter. web_pass.py:674 emits type="topic", tags=(…, "reference"); it becomes type="reference" with the tag dropped and _REFERENCE_CONCEPT_TAG (:169) deleted. The judgement-call paragraph at :85-92 goes with it.
• policy / capability — no code path has ever emitted them (RESEARCH M2). Resolving them as types is the deletion of the registry entries that expressed them as facets; nothing starts emitting them until a grain does. Saying this plainly matters: an implementer expecting an emission change will go looking for one that does not exist.
• methodology → playbook — the alias deletes. playbook also leaves the facet registry: it is already a type in the system-baseline lane (validator.py:246), and carrying one term in both vocabularies is the double-bookkeeping DR-141 diagnoses.

2.10 EffectiveOntology and the ontology artefacts (DR-027 as amended, F6)

EffectiveOntology survives with two dimensions. entity_types and relationship_types remain closed platform CVs — DR-027’s platform-repo half is explicitly unchanged by the S546 amendment, and lint_entity_relation_mentions is not in DR-141’s scope. concept_types leaves the dataclass; base_for_class deletes; base_only() returns the base pair; compose(overlay) unions the two remaining dimensions and ignores an overlay’s concept_types key.

_OVERLAY_DIMENSIONS (bundle_writer.py:817) keeps all three keys: the overlay schema is closed (OV-2), so removing concept_types would turn a valid client file into a validation failure. The key is still meaningful — it declares the client’s own vocabulary for the overlay echo, which bundle-graph.ts:338-350 reads as a styling signal. It simply no longer feeds a gate.

write_ontology_artefact emits overlay-only. _base_ontology_snapshot (:699-716) deletes; the payload becomes {"overlay": <mapping or null>}. Source: DR-027’s S546 amendment, executed here per its own text (“Implementation rides id-427’s wave”).

context.jsonld drops its concept_types dimension. iri_projection._DIMENSIONS (:76-80) loses that row; project_context (:227-245) projects entity_types + relationship_types only; the 5/12/10 counts in the slug() docstring (:138-139) correct to 12/10.

SUPERSEDED, in this paragraph only, by §6’s TQ-1 — RULED (S546), amended (S548), executed by {427.14}. project_context projects client-overlay terms only: the base entity/relationship half retires alongside the base namespace-prefix key, so a bundle composing no overlay emits {"@context": {}}. The _DIMENSIONS change above still stands; what changed is that its registers became an EXCLUSION filter rather than a mint list. The rest of §2.10 (EffectiveOntology, _OVERLAY_DIMENSIONS, write_ontology_artefact) is UNAFFECTED and landed as written at {427.11}.

Two riders {427.14} added that this section did not anticipate:

• _RESERVED_PREFIX_SLUGS drops "base", which was reserved solely to protect the prefix key that is now gone; keeping it would silently drop a client whose own overlay declares a term named base.
• The claim below that the overlay echo’s concept_types key is read by bundle-graph.ts:338-350 was measured WRONG at S550 — readBundleClassSignal only tested overlay for null-ness and never opened concept_types. {427.14} made the claim true rather than correcting it downward: that key is now the source of the re-sourced typeDeclaration border channel.

Why not re-derive the concept-type projection from the observed bundle instead. The classifier that made the projection meaningful was ALLOWED_CONCEPT_TYPES — the register that said which terms were base. With it gone there is no base/overlay distinction for that dimension, and minting a w3id.org/canonical/ontology IRI for an organically-minted label would assert exactly the registration DR-027-as-amended just retired ontology.json.base for. Carried as TQ-2 below.

2.11 Corpus census — making PI-2 mechanical

@dataclass(frozen=True)
class CorpusCensus:
    considered: tuple[tuple[str, int], ...] = ()   # ("source_documents", 42), ("q_a_pairs", 310)
    routed:     tuple[tuple[str, int], ...] = ()
    # unrouted derived per kind; unrouted_total is the property PI-8 keys on

Source gains census() -> CorpusCensus (both implementations). RunSummary gains census: CorpusCensus = CorpusCensus(). _render_run_bullets (bundle_writer.py:577) emits, unconditionally, including on a no-op run:

* **Run <ISO-ts> — Considered (2):** source_documents 42 (routed 42), q_a_pairs 310 (routed 310)

and, only when non-zero, an Unrouted (N) line. RunSummary.is_no_op (:562-570) gains or self.census.unrouted_total — a run that leaves knowledge unrouted is never a silent no-op, exactly the precedent failed already sets.

AMENDED 2026-08-10 by {427.13}‘s re-derivation (§3) — “unconditionally” is one case too broad, and the IMPLEMENTATION is the correct half. {427.9} shipped the Considered line guarded by if not census.considered: return []. The case this paragraph did not consider is a caller with no corpus at all — a direct write_bundle call, which is handed drafts rather than a Source and so takes no census. Emitting Considered (0) there would report a measurement nobody took, which is precisely the failure {427.7} named when it warned that its own empty Coverage “is not a measurement”; CorpusCensus()’s empty default means “no census taken”, not “a census that found nothing”.

The rule as amended: Considered is emitted on every run that took a census, including a no-op one — which is the guarantee this section actually wanted, since a no-op run over a real corpus is exactly the case the census exists to make speak. is_no_op is unaffected and landed as written.

Counts only — no uuids in log.md. BI-10 keeps record pointers to resource:/citations, and a run log is neither.

Rejected — aborting the run when unrouted > 0. The residual grain makes it structurally impossible, so a non-zero value means a grain regressed, not that the data is bad; DR-047’s degrade posture says report rather than destroy an otherwise-correct bundle. Silence is what is forbidden, and the census removes it. The hard guarantee lives in the behaviour test ({427.10}), where it belongs.

2.12 id-358 — the rename

ConceptKey.workspace_id → form_instance_id (l_records.py:219-236, and its __post_init__ check :285-290); ProposedChange.source_workspace_id → source_form_instance_id (git_sync.py:208-235); sync_bundle(source_workspace_ids=) → source_form_instance_ids= (:628,:653-656,:722-733); flow_def.py:617-638’s map variable follows. The BI-28 slot stays (owner, S546). The field has held a form_instances.id since {145.24}; the only thing wrong with it is the name, and the reason it was not renamed then — cross-file blast radius — is dissolved by this task rewriting those files anyway.

---

3. The audit — wired-sites disposition table (AC 5)

Substrate §2’s blast-radius list, each site dispositioned. Changed / kept / deleted / UNDECIDABLE. Sites the substrate listed that measurement moved are flagged [C]* with the challenge that moved them.

Producer — scripts/cocoindex_pipeline/producer/

Site	Disposition
validator.py:116-118 ALLOWED_CONCEPT_TYPES	DELETED (§2.9)
validator.py:120-157 RECOGNISED_FACET_TAGS, FACET_TAG_ALIASES	DELETED (§2.9) [C3]
validator.py:160-175 canonical_facet_tag, normalise_facet_tags	DELETED — no production caller (M3)
validator.py:242-248 _CLASS_CONCEPT_TYPES	DELETED — owner ruling (iii)
validator.py:263-331 EffectiveOntology	CHANGED — two dimensions; base_for_class deleted (§2.10)
validator.py:406-420 check_type_membership	CHANGED → check_type_shape (§2.8)
validator.py:596-617 check_concept / validate_concept	CHANGED — calls the shape check; stops threading effective_ontology into it
l_records.py:115-117 CONCEPT_TYPES	DELETED
l_records.py:152-168 contextvars overlay widener	DELETED
l_records.py:219-236, 285-290 ConceptKey.workspace_id	CHANGED — renamed form_instance_id (id-358)
l_records.py:260-290 ConceptKey.__post_init__	CHANGED — shape + BI-3 + non-empty grain
l_records.py:340-351 local Source protocol	MOVED to sources/base.py; gains census() (id-362 F1)
l_records.py:676 _slugify	KEPT, wrapped by mint_concept_slug (§2.6)
l_records.py:736-753 list_concepts	CHANGED — registry iteration + coverage union + residual last
l_records.py:755-821 _list_feeder_concepts	KEPT, re-described — client-declared routing grain; config gains directory (§2.7)
l_records.py:823-984 the six grain methods	CHANGED — each returns GrainEnumeration; each declares its directory
l_records.py:993-1017 read_concept five-way dispatch	CHANGED — grain-keyed (§1)
l_records.py:1056-1085 _source_documents_for_key	CHANGED — takes patterns; stops sniffing concept_type; its topic-branch ValueError deletes
l_records.py:1185-1210 sample_rows	CHANGED — grain-keyed
bundle_writer.py:158 validator import	CHANGED
bundle_writer.py:224-284 won-bid redirect, bundle_write_path, bundle_write_path_for_key	DELETED (§2.5)
bundle_writer.py:528-570 RunSummary	CHANGED — census; is_no_op accounts for unrouted (§2.11)
bundle_writer.py:577-620 _render_run_bullets	CHANGED — unconditional Considered line
bundle_writer.py:699-716 _base_ontology_snapshot	DELETED — DR-027 S546 amendment
bundle_writer.py:745-747, 787-812 overlay class gate	KEPT — DR-054/DR-079, a different live requirement
bundle_writer.py:918 _validate_concept_feeder_schema	CHANGED — accepts directory; rejects index/log (§2.6)
bundle_writer.py:1025-1054 write_ontology_artefact	CHANGED — overlay-only
bundle_writer.py:1266-1282 bundle-class → effective ontology	CHANGED — no concept-type dimension; the internal_dev fail-loud deletes (its requirement was “no ratified type set yet”)
bundle_writer.py:394-519 IndexTheme / build_index_themes / theme_config / unthemed_heading	RETIRED BY id-429 (D3) — not dispositioned here; shared-file coordination only
iri_projection.py:57-62, 76-80, 227-245	CHANGED — concept_types dimension dropped (§2.10)
iri_projection.py:138-139 slug() docstring 5/12/10	CHANGED — docstring counts correct to 12/10
enrich.py:246 imports from sources.l_records	CHANGED — imports from sources/base.py (id-362 F1)
enrich.py:328-341 _qa_pairs_anchor	CHANGED — grain-keyed, not type-keyed
enrich.py:344-353 _resource_from_raw	KEPT mechanically; its output relocates to sources[] — id-426 owns that move
enrich.py:380-395 _annotate_raw_with_anchors	KEPT — rename ripple only
enrich.py:847-880 enrich_concept	CHANGED — protocol-typed; version= bump (§5) — see the S550 correction under this table
enrich.py:938-951 frontmatter assembly	CHANGED — template path bypasses; confidence: no-content reachable
web_pass.py:85-92 judgement-call docstring, :169 _REFERENCE_CONCEPT_TAG, :670-674	CHANGED — type="reference", tag dropped (§2.9)
flow_def.py:138, 519-533 _REPO_DOCS_BUNDLE_CLASSES	KEPT [C1] — source selection (bundle-doctrine’s two paths), not type gating
flow_def.py:617-638 BI-28 provenance map	CHANGED — rename ripple; keys on rel_path once the redirect goes. Residual stamping: TQ-3 / PQ-3
git_sync.py:208-235, 628-656, 712-733	CHANGED — id-358 rename (§2.12)

Sources — scripts/cocoindex_pipeline/sources/

Site	Disposition
repo_docs.py:105-119 SYSTEM_BASELINE_CONCEPT_TYPES	DELETED — with _CLASS_CONCEPT_TYPES
repo_docs.py:122-201 RepoConceptKey	CHANGED — shape check replaces membership; converges on the shared ConceptKey (id-362 F1)
repo_docs.py:234-242 duplicate Source protocol	DELETED [M14] — imports the shared one
repo_docs.py:549-680 grains at :584/:594/:628	CHANGED — declare directories + coverage; census() added — see the S550 correction under this table

Correction (2026-08-10, S550, from id-427 {427.16}) — two rows of this table asserted an end state as a change

Both rows above were read as delivered and neither was. The S548 independent spec-compliance audit measured them; {427.16} executed the fix. Recorded here because the table itself was the mechanism — “the disposition table asserted the end state as a change, so both the implementer and the ledger read it as delivered” (tasks/id-362.md, S548).

• enrich_concept — “CHANGED — protocol-typed” described a state that ALREADY HELD at the base SHA. The function was byte-identical across the whole wave apart from version=3; it was typed key: ConceptKey, which is precisely NOT protocol-typed. It is now key: ConceptKeyLike. And the row understates the work even as an instruction: typing alone leaves the call raising AttributeError on a RepoConceptRaw, because §3 nowhere notes that ConceptRaw and RepoConceptRaw share no field. The row should have read “CHANGED — protocol-typed AND the anchor-annotation path taught the artefact raw”.
• repo_docs.py grains — “declare directories + coverage; census() added” was HALF done. {427.9} added Coverage and census() and disclosed the rest as an open gap; the registry and the declared directories landed only at {427.16}. A row conjoining four deliverables with + cannot record a partial landing, which is how the disclosed gap went unowned between {427.9} and the S548 audit.

Every line number in both rows is stale (measured S550): the hardcoded pillar tuple sat at repo_docs.py:610-613, the directory literals at :707/:709/:728, and enrich_concept at :932-938. Line numbers in this table are pre-{427.4} and should be treated as locators to re-derive, never as coordinates.

A third site this table never named: flow_def.run_producer_flow’s BI-28 provenance map read key.form_instance_id directly, crashing every staging system_baseline/internal_dev run. §3’s flow_def.py:617-638 row calls the BI-28 map “CHANGED — rename ripple; keys on rel_path once the redirect goes” — true as far as it goes, and blind to the fact that the line runs for both key models.

RE-DERIVATION (2026-08-10, S550, id-427 {427.13} — AC 5) — every row measured against 20b644468

Method, and why it is not a re-read. This table had been measured wrong four times across two sessions, every time by walking it as written. So no row below was settled from the row above it. Each was settled by measuring the symbol at the base SHA and diffing its site against the pre-wave commit 2b21fdc05 (the parent of {427.4}, 932b3b784) — because enrich_concept taught that “CHANGED” can describe a state that already held, and only a diff against the pre-wave tree distinguishes a change from a description.

Coverage: 56 of 56 table rows re-derived (41 Producer, 4 Sources, 7 TypeScript, 4 Bundle artefacts), plus the Tests block and the “Explicitly not swept” block — 58 dispositions, none left blank. What this pass did NOT do: it did not run the producer, execute any SQL, or open a database; it did not re-read the vendored OKF SPEC; it did not count test assertions (the substrate’s caveat still stands); and it did not verify the four rows the owner ruled flag-and-route beyond establishing their current state. GitNexus’s index was measured stale against this SHA — it surfaced test_an_overlay_declaring_concept_types_stays_schema_valid_and_ composes_nothing, a test that has since been inverted — so every verdict here rests on grep + git diff + direct reads, not on the code-intelligence graph.

Seven rows were wrong. Each is stated with what it asserted.

1. bundle_writer.py:577-620 _render_run_bullets — “CHANGED — unconditional Considered line” is FALSE, and the row is not where the error starts. The line is explicitly CONDITIONAL: _render_census_bullets opens if not census.considered: return [], and its docstring gives the reason — a run with no census (a direct write_bundle call, which never saw a corpus) emits neither line, because Considered (0) there “would report a measurement nobody took, which is the error {427.7} named when it warned that its own empty Coverage ‘is not a measurement’”. What holds unconditionally is narrower: Considered is emitted on every run that took a census, including a no-op one.

   The row faithfully restates §2.11, so §2.11 is the stale carrier. §2.11 specifies that _render_run_bullets emits “unconditionally, including on a no-op run”. {427.9} deliberately departed from that when it met the case §2.11 had not considered — a caller with no corpus at all — and the departure is right: the implementation distinguishes “measured, and found nothing” from “never measured”, which the spec’s blanket rule cannot. This is an undocumented spec departure where the CODE is correct and the SPEC was never updated, which is the reverse of this table’s other six errors. §2.11 is annotated accordingly.

2. enrich.py:380-395 _annotate_raw_with_anchors — “KEPT — rename ripple only” is FALSE. It gained an entire artefact branch (text-discriminated, minting through the same _mint ledger so BI-17 holds) and is the change that closed id-362 F1 leg 2. This is the SAME defect the S550 correction block above already diagnosed one row later: that block corrected enrich_concept’s row to read “protocol-typed AND the anchor-annotation path taught the artefact raw” — and left the anchor-annotation path’s OWN row still saying “rename ripple only”. The correction named the work and did not follow it to the row that owned it.

3. enrich.py:328-341 _qa_pairs_anchor — “CHANGED — grain-keyed, not type-keyed” is wrong at BOTH ends. It is not grain-keyed now: it reads getattr(key, "scope_tag", None), a locator presence test that consults no grain registry. And it was not type-keyed before: pre-wave it read key.scope_tag is not None, the same presence test, unguarded. The row describes a transition between two states that never held. The function DID change ({427.16}: ConceptKeyLike + getattr, so a RepoConceptKey with no scope_tag gets the honest None), so the disposition word survives and its stated mechanism does not.

4. Tests — “All CHANGED” is FALSE for one of the ten named modules. scripts/tests/test_producer_frontmatter.py is byte-identical across the entire wave (git diff 2b21fdc05..20b644468 returns empty for it). The other nine did change. This is the enrich_concept failure mode in the test block: a blanket disposition that reads as delivered for a module nothing touched.

5. l_records.py:755-821 _list_feeder_concepts — “KEPT, re-described” understates a rename and a re-signature. It is now _list_feeder_grain(spec: GrainSpec) -> GrainEnumeration; pre-wave it was _list_feeder_concepts() -> list[ConceptKey] and carried the grain == "entity_mention" dispatch itself. That dispatch and its ValueError moved out to the registry-build path (_feeder_grains). The row’s substantive claim — config gains directory — is delivered (directory=grain_config.get("directory") or concept_type).

6. l_records.py:676 _slugify — “KEPT, wrapped by mint_concept_slug” is true but mislocates it. Both functions live in sources/base.py now, not l_records.py. The same applies to three neighbouring rows: ConceptKey.workspace_id, ConceptKey.__post_init__ and the local Source protocol are all dispositioned against l_records.py line ranges, and all three symbols now live in sources/base.py. Only the Source row says so.

7. Bundle artefacts — showcase ontology.json “REGENERATED overlay-only on the next run” has NOT landed, and cannot have. The on-disk artefact at okf-bundles/canonical-okf-showcase/ontology.json still carries a base key (with entity_types + relationship_types; no concept_types). The row is a forecast contingent on a producer run that has not happened, which is a legitimate state — but it is PENDING, not delivered, and a table walked for “landed” would have ticked it.

Two more sites this table never named (the pattern the S550 block opened with flow_def):

• bundle_writer._render_census_bullets — net-new at {427.9}, and the function that actually renders the census. §3 dispositions only its caller, which is how correction 1 above went unnoticed.
• enrich._samples_source_documents — net-new, and the genuinely grain-keyed function (getattr(grain_for(key), "sample_kind", "")). It is the mechanism correction 3’s row wrongly attributed to _qa_pairs_anchor.

The remaining 49 rows are CONFIRMED as written — each measured, not read across. In summary, by disposition:

Disposition	Rows	Verification executed
DELETED (11 symbols across 9 rows)	ALLOWED_CONCEPT_TYPES, RECOGNISED_FACET_TAGS, FACET_TAG_ALIASES, canonical_facet_tag, normalise_facet_tags, _CLASS_CONCEPT_TYPES, CONCEPT_TYPES, contextvars overlay widener, won-bid redirect + bundle_write_path + bundle_write_path_for_key, _base_ontology_snapshot, SYSTEM_BASELINE_CONCEPT_TYPES, duplicate Source protocol, CONCEPT_TYPE_VALUES	Each confirmed a REAL definition at 2b21fdc05 (git grep at that rev) and absent at 20b644468 — surviving only as prose in comments, and as assert not hasattr(...) guards in test_producer_validator.py:697-700/799-801 and test_producer_bundle_writer.py:1535. Absence alone was NOT taken as evidence; the pre-wave definition is what makes each a deletion rather than a thing that never existed.
CHANGED (confirmed)	EffectiveOntology (3 fields → 2; base_for_class gone), check_type_membership→check_type_shape (takes no effective ontology), check_concept/validate_concept (calls the shape check; stops threading effective_ontology into it — still threads it into lint_entity_relation_mentions, which is what the row means), ConceptKey.workspace_id→form_instance_id, __post_init__ (shape + BI-3 + non-empty grain), list_concepts (registry iteration, coverage union, runs_last ordering), the six grain methods, read_concept (registry lookup), _source_documents_for_key, sample_rows, validator import, RunSummary (census field; is_no_op reads census.unrouted_total), _validate_concept_feeder_schema (accepts directory, refuses index/log), write_ontology_artefact (payload is exactly {"overlay": …}), bundle-class→effective ontology, iri_projection dimensions + slug() docstring counts, enrich.py imports, enrich_concept, frontmatter assembly, web_pass (_REFERENCE_CONCEPT_TYPE = "reference"; no _REFERENCE_CONCEPT_TAG), flow_def BI-28 map, git_sync (id-358 rename), RepoConceptKey, repo_docs grains, concept-schema.ts parity docstring	Symbol located at HEAD, then git diff 2b21fdc05..20b644468 over its site to prove the change is real rather than a description of the status quo ante. Two carry riders: _source_documents_for_key is now _source_documents_by_patterns(patterns) — the key parameter is dropped ENTIRELY, not merely supplemented, and its own docstring records the deleted topic-branch ValueError; bundle-class → effective ontology no longer reads bundle_class at all (EffectiveOntology.compose(overlay) is the single path), so the internal_dev fail-loud is deleted rather than relocated. The six grain methods return GrainEnumeration as stated, but their directories are declared on GrainSpec in _BUILTIN_GRAINS, not in the methods.
KEPT (confirmed, with its live requirement named)	overlay class gate (requirement: only client_business may compose a client overlay; source: DR-054/DR-079, OV-10 — live, asserted at bundle_writer.py:1722 and its own docstring calls itself UNCHANGED), flow_def._REPO_DOCS_BUNDLE_CLASSES [C1] (requirement: source selection between bundle-doctrine’s two paths; source: ID-163 PC-2 — live, read at :559), concept-schema.ts type: z.string().min(1) (already correct pre-wave), KNOWN_TYPES as legend (extended past the base 5; 018fc2270 made the domain-tokens guard enumerate it rather than a copy), bundle-graph.ts type derivation, bundle-graph.ts ontology.json overlay read (readBundleClassSignal reads the key’s PRESENCE only), the five type-named directories [C2] (all five present and unmoved on disk)	Each verdict states the requirement and its current source, per the requirement-first rule. A “KEPT” row with no nameable live requirement would have been UNDECIDABLE; none was.
UNSWEPT / NOT DISPOSITIONED (unchanged status, restated)	parse-index.ts/parse-log.ts; IndexTheme/build_index_themes/theme_config/unthemed_heading; enrich._resource_from_raw; CONFORMANCE.md rows	See the routing block below. The id-429 theme row is worth one measured note: it says “RETIRED BY id-429 (D3) — not dispositioned here”, and that retirement has since landed (9037022e3) — IndexTheme is renamed, theme_config and unthemed_heading are gone. Still not dispositioned by this task; recorded so the row is not read as outstanding.

No row in this table is UNDECIDABLE. Every one had a nameable requirement with a current source, or a measurable presence/absence at two SHAs. The four TQ-* UNDECIDABLEs in §6 are questions about the DESIGN, not about a row’s disposition, and they are carried verbatim into tasks/id-427.md by {427.13} rather than resolved here.

Flagged and routed, not fixed — the owner’s S550 directive is that {427.13} must not complete work that an in-flight task will change. Each carries its owning task id:

Site	Owner	State measured at 20b644468
showcase CONFORMANCE.md superset 1	id-431	AC 6 is discharged at the register, not in the file — DR-019’s S545/S546 amendment already reads “Divergence 1 — the closed, validator-enforced type taxonomy — is WITHDRAWN”, and authority for a conformance posture lives on the register. DR-019 itself records the CONFORMANCE.md house rule as “re-opened with id-431”, whose entire goal is deciding whether a bundle should carry the file at all. Editing a file id-431 may delete is the churn the directive rules out. One brief-supplied ground was measured FALSE and is corrected here: the bundle directory okf-bundles/canonical-okf-showcase/ IS a git repository (git rev-parse --is-inside-work-tree → true, history back through bc5a80a) and CONFORMANCE.md IS tracked (git ls-files returns it). The other two grounds stand and are sufficient.
showcase CONFORMANCE.md superset 3 (ontology.json)	id-430 / id-431	Unchanged — id-430 rules the artefact, id-431 the file.
enrich._resource_from_raw	id-426	Mechanically KEPT as the row says; re-typed to (key: ConceptKeyLike, raw: Any). Its output’s relocation to sources[] is id-426’s.
bundle_writer IndexTheme / build_index_themes / theme_config / unthemed_heading	id-429	Retirement HAS landed (9037022e3) — see the table above.
Anything v0.2, okf_version, frontmatter shape, sources[]	id-426 (producer) / id-439 (consumers)	Both doing; untouched here.
enrich._seed_user_message’s “backing L-records” prompt text	{427.16} flagged it	Changing prompt text is a DR-060 drafting-config change requiring a version= bump, and this wave has already spent its one bump (version=3, TECH §5). Flag only.
__tests__/lib/okf/bundle-graph.test.ts fixtures shipping concept_types inside ontology.json	id-439	4 sites measured — :377, :427, :451, :463. Left deliberately by the S550 overlay retirement; the consumer-side parser work is id-439’s.
_RETIRED_OVERLAY_DIMENSIONS tombstone branch	future subtask	Its stated retirement condition is now MET (see §OV-2’s amendment); retiring it is a behaviour change, so {427.13} recorded the discharge in the docstring and left the branch standing.

TypeScript

Site	Disposition
lib/ontology/concept-schema.ts:70 CONCEPT_TYPE_VALUES	DELETED [M4]
lib/ontology/concept-schema.ts:25-46 type-parity docstring	CHANGED — restated against DR-141
lib/ontology/concept-schema.ts:136 type: z.string().min(1)	KEPT — already correct
lib/okf/concept-type-tokens.ts:38-48 KNOWN_TYPES	KEPT AS LEGEND, extended (§2.9)
lib/okf/bundle-graph.ts:515 type derivation	KEPT — already derives from the bundle
lib/okf/bundle-graph.ts:338-350 ontology.json overlay read	KEPT — F6 keeps overlay
lib/okf/parse-index.ts, parse-log.ts	UNSWEPT — the census adds a log.md line whose parser this spec did not read; verified in {427.9}, not assumed

Tests

test_producer_validator, test_l_records_source, test_producer_bundle_writer, test_producer_iri_projection, test_producer_enrich, test_producer_flow_def, test_producer_frontmatter, test_producer_bi28_bidoutcome_proposal; __tests__/lib/ontology/concept-schema.test.ts, concept-type-tokens.test.ts. All CHANGED. Assertion counts unswept (substrate’s own caveat stands); the four assertions grepped this pass are test_producer_validator.py:449-493 (facet registry) and concept-schema.test.ts:73 (CONCEPT_TYPE_VALUES), all of which delete with their subjects.

Bundle artefacts

Site	Disposition
showcase CONFORMANCE.md:35 superset 1	CHANGED — withdrawn in step (AC 6); coordinate id-431
showcase CONFORMANCE.md:41 superset 3 (ontology.json)	NOT THIS TASK — id-430 ruled the artefact, id-431 rules the file
showcase ontology.json	REGENERATED overlay-only on the next run
the five type-named directories	KEPT UNCHANGED [C2] — they are grain constants, not a type materialisation

Explicitly not swept

.claude/worktrees/agent-* stale producer copies; supabase/migrations; specs/id-132-* DR-141 propagation. All three carried forward from the substrate, all three still open.

---

4. Migration and compatibility — the first post-inversion run

Measured surface: 19 concept files across five directories plus case-studies/won-bid/ (RESEARCH M11).

Question	Answer
Do the five directories move?	No. They are grain constants and the grains keep them (§2.4).
Does any existing file move?	No. RunSummary.moved stays empty. The won-bid change is to identity, not to the physical path (§2.5).
Do the 19 concepts re-draft?	Yes, once — ConceptKey gains grain and loses/renames workspace_id, and cocoindex fingerprints every field unconditionally (l_records.py:135-145). This is a memo miss, not a move.
Does any concept’s type change?	No for all 19 — each grain keeps the label it already emits. reference concepts change type, and none is shipped (M11: no references/ directory exists).
What is new on disk?	documents/, and questionnaire-responses/ / unattributed-answers/ if the corpus has residue. Every new file lands in RunSummary.added, which is the correct report.
What changes in ontology.json?	base disappears; overlay: null stays.
What changes in context.jsonld?	No published instance exists to change (M11 / substrate §4).
Does CONFORMANCE.md change?	It is hand-authored and producer-preserved (bundle_writer.py:186-202); superset 1’s withdrawal is a manual edit in {427.13}.

RunSummary.moved semantics are unchanged and stay caller-supplied (bundle_writer.py:1172-1180 — a move cannot be inferred from a flat path diff). Nothing in this task supplies one.

---

5. Memo, fingerprint and DR-060

ConceptKey is the @coco.fn(memo=True) key and every field fingerprints unconditionally (enrich.py:863-871, l_records.py:135-145). This task changes the field set twice (grain added, workspace_id renamed) and changes the output shape (template drafts, confidence: no-content). Per DR-060 that is a manual version= bump on enrich_concept, recorded in the bundle log.md at the producer run — one bump for the whole wave, not one per subtask. grain is added immediately after concept_type and the “keep new fields last for positional compatibility” convention (l_records.py:253-258) is explicitly retired for this wave: every construction site is being rewritten anyway, and a silent positional shift is worse than a compile-time break.

---

6. UNDECIDABLE (carried verbatim)

TQ-1 — RULED (S546), amended (S548), EXECUTED by {427.14} (S550). No longer open. “What live document requires the bundle to declare a vocabulary at all, now that ontology.json.base retires and context.jsonld loses its concept-type dimension — or is the remaining entity_types/relationship_types projection surviving on DR-027’s platform-repo half, which says nothing about what is asserted to consumers?” — DR-027’s S546 amendment is explicit that base CVs “are simply no longer asserted to bundle consumers”, yet context.jsonld continues to assert two of the three dimensions to exactly those consumers. This spec keeps them because DR-141 scopes only the concept-type vocabulary and unasked-for retirement is not this task’s to take. The question belongs to id-430’s ruling family, and it is not closed.

Resolution. RULED S546 (owner-delegated): overlay-driven emission only; the base-only projection retires. The ruling’s stated ground — “no consumer” — was measured FALSE at S548 (bundle-graph.ts rendered context.jsonld into a node border colour) and the ruling was re-grounded on DR-027’s other and stronger ground; see DR-027’s S548 amendment, which exists so the “no consumer” claim is never re-derived. Executed S550 by {427.14}, which also repaired the channel the artefact fed (re-sourced onto ontology.json’s overlay.concept_types) rather than leaving it vestigial.

TQ-1a — still OPEN, carried verbatim in tasks/id-427.md {427.14}. Which of the ruling’s two emission readings is meant (artefact not written at all on a no-overlay run, vs always written carrying only overlay terms). {427.14} shipped the second as a REVERSIBLE DEFAULT, not as an answer: it could not name a requirement whose current source discriminates them, since both satisfy “declare the client’s overlay vocabulary” vacuously when there is no overlay. OV-10 does not settle it by analogy — that precedent rests on a consumer that distinguishes present-and-null from absent, and no reader distinguishes an empty @context from a missing file.

TQ-2. “If a future RDF/JSON-LD export needs a dereferenceable rdf:type IRI per concept, what mints it once concept-type labels are unregistered — and does that requirement exist today, with a current source?” — DR-082 ratifies the namespace authority; nothing found this pass names an export that consumes it. §2.10 drops the projection on the ground that minting an IRI for an unregistered label asserts registration. If the requirement is later shown live, the honest re-entry is a per-concept IRI derived from the concept’s own bundle path, not a vocabulary declaration.

TQ-3. “Should a residual questionnaire_response concept carry BI-28 form-instance provenance the way a won-bid proposal does, and what is the current source for the rule that decides?” — BI-28’s “never an unattributed proposal” argues yes; the owner’s S546 caveat that “the entire procurement functionality is under review as it’s clear that there is confusion and duplication currently” argues for touching nothing. Not stamped by this spec. Mirrors PRODUCT PQ-3.

TQ-4 (carried from PRODUCT PQ-2). “On the day id-422’s publication gate lands and an empty-scope_tag published pair becomes impossible, is the residual cascade’s defensive branch retired or kept permanently?” — the S546 ruling says “enumerated defensively until that gate lands” and stops there. Kept behind one grain-registry entry so either answer is a one-line change.

7. What this spec did not cover

• No producer run and no database query. Every SQL shape in §2.1 is specified, not executed; the anti-join plans are unmeasured against a real corpus.
• parse-index.ts / parse-log.ts internals — the census adds a log.md line and this spec did not read the log parser. {427.9} verifies rather than assumes.
• Test bodies beyond the four assertions in RESEARCH M3/M4.
• The vendored OKF SPEC was not re-read; id-426’s S546 delta note is the sole authority for every v0.2 claim here.
• .claude/worktrees/agent-*, specs/id-132-* DR-141 propagation — unswept, carried from the substrate.
• id-429’s own surfaces — IndexTheme/build_index_themes/theme_config/ unthemed_heading are named in the disposition table as retired by id-429 and are deliberately not dispositioned here, though both tasks edit bundle_writer.py.

S546 post-authoring rulings — the five carried questions (owner-delegated)

The owner reviewed the spec report and delegated these five to staff-engineer judgment (S546). Rulings, each with its ground:

• PQ-1 — RULED: document existence is disclosable; un-admitted content is not. The bundle is the client’s own repo, and every existing concept already cites source_documents by pointer — a residual concept adds no new disclosure class. Rider binding the residual-template subtask: the template renders only admission-surface metadata (title/path/dates/counts — the same fields citations already expose) and never quotes content that has not passed the knowledge-admission gate (R2). Onward-sharing risk is governed by the client’s own sharing decision, exactly as it already is for every concept.
• PQ-2 / TQ-4 — RULED: retire the defensive branch when id-422’s gate lands. The invariant moves to the DB CHECK (enforced by construction at the write layer); an unreachable branch kept as belt-and-braces is the exact DR-139 pattern this programme retires. Routed: an AC added to id-422 — the gate-landing change removes the branch and its test asserts the CHECK instead. The residual complement needs no special case; anti-join coverage is unaffected.
• PQ-3 / TQ-3 — RULED: yes — carry form_instance_id provenance on residual questionnaire_response concepts. The attribution key IS the form instance (cascade step 2), so the value is already in hand; BI-28 is ruled live (QC3-A); two grains attributing by the same key carry the same provenance shape.
• TQ-1 — RULED: context.jsonld follows ontology.json’s F6 fate — overlay-driven emission only; the base-only projection retires. The same principle applied uniformly: a consumer-asserted vocabulary artefact with no consumer and no requirement source is not emitted. The entity/relationship vocabulary remains platform-internal (DR-027’s unchanged half). Supersedes {132.44}‘s base-only-plus-advisory behaviour (pre-launch schema/behaviour change per DR-139). Folds into PLAN’s F6 subtask.
• TQ-2 — RULED: no live requirement; defer, documenting the door. RDF export is a benefit note in DR-027, not a requirement with a current source. If it materialises, a label type mints a deterministic IRI via the existing slug() under the DR-082 namespace — the mechanism needs no registry. Nothing is built now.

Also executed: the DR-141 rider (coverage ≥1, not partition) — recorded on the register.