Canon 0.5.0: uses_provisions so P stays purchased platform

A provision can name which other provision satisfies a catalog
dependency. Relation vocabulary is depends_on / may_use only.
Quoted sits below the invoiced/measured peer pair for propagation.

CHANGELOG.md

@@ -15,6 +15,38 @@ semantics.
 
 ---
 
+## 0.5.0 — 2026-08-15
+
+### Added — Provision-to-provision relationships (`uses_provisions`)
+
+**Change.** A `CapabilityProvision` may name the provisions it relies on
+under `uses_provisions`. Each entry is a catalog capability, a consumer
+landscape `provider` reference, and a `relation` of `depends_on` or
+`may_use` — the existing capability-to-capability vocabulary, not a new
+one. A `depends_on` entry MUST be declared between those capabilities in
+the catalog; a `may_use` entry SHOULD. Using another capability is not
+consuming class `P` (CAP-R11).
+
+`data.backup` gains catalog `may_use: security.secrets`, from the
+resource-control restatement. Maturity is not derived or capped across
+the dependency.
+
+**Rationale.** The catalog can say `data.backup depends_on data.object`;
+a provision could not say *which* object provision satisfies that for
+it. The leftover was a `consumes:P` row for OpenBao credential custody —
+the wrong kind, and the start of `P` widening again after finding C
+narrowed it. Filed as `demand/ProvisionRelationships.md`.
+
+**Breaking:** no capability id changed. New optional field on the
+still-proposed provision record. Consumers who wrote `relation: uses`
+should spell it `may_use`.
+
+**Records.** `ADHOC-2026-08-15` T03; demand accepted.
+
+**Validation.** `make validate` 0 errors / 0 warnings; `make test` 22 passed.
+
+---
+
 ## 0.4.0 — 2026-08-15
 
 ### Changed — Joinable consumer proof counts for ITC-CAP promotion
@@ -41,8 +73,8 @@ joinable provision that already fails or passes against this catalog.
 
 **Change.** ITC-GOV gains `EvidenceBasis`: a closed vocabulary for how a
 quantity was obtained, in **tiers** rather than a total order
-(`evidenced`: invoiced / measured / quoted are peers; then `projected`;
-then `judgement`: estimated / assumed; then `unknown`). A derived value
+(`evidenced`: invoiced / measured / quoted; then `projected`; then
+`judgement`: estimated / assumed; then `unknown`). A derived value
 resolves to the weakest *tier* among its inputs. Unknown-is-not-zero.
 Optional `proxy_for` / `observed_at`. Decision grades (`evidenced` /
 `projected` / `indicative` / `insufficient`). Catalog: