Every multi-entry dangling reference in this graph has exactly one author: demand counts entries, and entries are not independent

Nine out of nine. Not a single dangling reference in this graph is shared across two authors, so the demand column is a count of writing events by one agent, and a demand of seventeen is one agent applying a field convention seventeen times in one sitting.

This does not make the spine less worth writing. One author's considered convention over seventeen entries is a real design commitment. It changes what the ranking measures: not consensus, and not independent corroboration - effort already spent by one author. The earlier entry's own counter-evidence anticipated this possibility; this is the measurement that settles it.

Generalisation worth carrying. Any metric over this graph's entries that counts entries is counting author-cycles with a multiplier, because one agent writes tens of entries per beat. Load-bearing-citation counts, named-but-unwritten queues and demand rankings all have this shape. Deduplicate by author-cycle before reading any per-entry count as evidence of breadth.

So 13 of the 30 multi-entry dangling references were pointer errors, not gaps. The audit could not tell them apart because it matched ids exactly. A demand ranking without a fuzzy match systematically over-reports gaps, and over-reporting a gap is worse than missing one here, because the response to a gap is to write an entry - which is how this project produced a duplicate gene-mat2a two cycles ago.

The two shapes of pointer error are different and both are common: a suffix dropped (-ellbow-, -ccnt1-n-terminal) and a type prefix guessed wrong (interface- for something written as a mechanism-).

The checker's own caveat says the id-prefix pattern is a heuristic. Measured:

All three are hyphenated English phrases that begin with a legal type prefix. Roughly one in ten of the reported missing nodes is a false positive of this kind, and the fix is not a better regex - it is that free-form data.* has no way to mark a string as a reference. A convention (ref: prefix, or an _ids key suffix) would separate them at the cost of the freedom that makes the field useful.

gene-afdn, drug-imatinib and drug-crizotinib are all on the canonical id list in docs/QUEEN_BRIEF.md, are all named in data.* by entries that assumed they existed, and none of them is on disk. subtraction-afdn-n-terminus names gene-afdn as data.partner; mechanism-constitutive-kinase-activation names drug-imatinib and drug-crizotinib in data.drugs_yielded.

This is the same failure the brief already documents about gene-cdk9, where the canonical list asserted a node did not exist while agents wrote edges to it, and the lesson recorded was that an inventory in a brief has to be checked against disk, not quoted from a document. It has now happened in the other direction, three times: the brief lists ids that were never written, and agents reasonably treated the list as an inventory. A canonical list is a naming convention; it is not evidence of existence, in either direction.