Is cessation already representable?

The standing rule requires exhausting composition before accepting architectural pressure. Applied to temporal extent.

Verdict: already representable — and blocked by a canonicalization defect, not by the ontology.

---

The six encodings

Each uses only existing concepts. None requires a new primitive.

Phenomenon	Encoding	Loses truth?
employee leaves	condition: unavailable on the actor	no — the person still exists, and is unavailable
service decommissioned	condition: unavailable on the resource	no
contract expires	condition: unavailable on the agreement	no
metabolite consumed	measure: concentration = 0	no — and it is more precise than cessation
package delivered	placement changes to the new place	no — the package persists, its location changed
connection terminated	condition: unreachable	no

The metabolite case is the strongest. concentration = 0 is better than "ceases to exist": it preserves the missing-versus-negative distinction the whole provenance apparatus rests on. Absence of a fact means nobody measured; a fact with value 0 means someone measured and found none. Modelling cessation as subject-deletion would destroy that distinction.

So the ontology does not need cessation. It needs magnitudes that can change.

---

Why the composition fails anyway

It fails at collapse, not in the vocabulary. Measured with a correctly typed fixture:

identity: measure:concentration:resource:cell/glucose
same identity for both observations? true

early first  -> value=1  validAt=00:10   ← the LATE timestamp
late  first  -> value=0  validAt=00:10   ← with the EARLY value

Two defects, in one function:

1. collapse is order-dependent

The same two facts produce a different world depending on array order. A canonical model whose output depends on input ordering is not canonical.

2. It fabricates a fact that was never observed

collapse advances validAt to the latest while keeping the first-encountered value. The surviving fact therefore asserts "at 00:10, concentration was 1 mM" — when the observation at 00:10 said 0 mM. That is not information loss; it is a fact nobody reported.

The root

factIdentity for a measure is measure:${metric}:${subject} — time-invariant identity for a time-varying quantity. Two measurements of the same quantity at different times therefore collapse as duplicates.

For condition the assumption holds: a condition is present or absent, and two observers reporting it are reporting one fact. For measure and placement it does not: the same subject and metric at two times is two observations of a changing quantity, not one fact seen twice.

The vocabulary carries validAt on every fact. The identity function discards it. The information is present and canonicalization throws it away — which is the same failure shape as every other defect this audit has found.

---

Classification

Candidate	Verdict
already representable	yes — by composition
composition	condition: unavailable, or measure = 0, or a placement change
architectural pressure	no
false lead	no — the phenomenon is real, the diagnosis was wrong

Temporal extent is not architectural pressure. Window remains unused because nothing needs it: a bounded existence is expressible as a magnitude falling to zero or a condition becoming true. What is needed is that a later observation of the same quantity supersede an earlier one, which is a defect in collapse and factIdentity.

Hypothesis 3 confirmed — the mapping into the vocabulary was wrong. The ontology is not implicated.

---

A retraction

I first ran this with any[] fixtures and reported that factIdentity produced [object Object], concluding that every measure of a metric collapsed regardless of subject. That was my error. Fact.subject is typed Urn, a string; my fixture passed a SubjectRef object. Using any[] disabled exactly the type check that would have caught it.

Same shape as every measurement error in this project: the instrument was wrong, and it was wrong because I had switched off the thing that would have told me.

---

Why this defect has never been observed

core has no consumers. lib/world uses its own mergeFacts (later substrate wins — correct), not core.collapse. So the defect is latent: it lives in the most heavily verified module in the repository, and the verification never exercised a changing quantity, only duplicate observations of a static one.

That closes a loop with the platform-boundary finding. core's zero-consumer status is not merely an adoption gap — it means core's defects go undetected, because nothing pushes real time-varying data through it.