Expand description
The production PermissionMap (#676): the S3 scope ↔ S5 TE-rule
vocabulary, injective and exact.
§Shape (and why)
A TeRule is a concrete numeric triple (subject_type, object_type, action) — a wildcard is unrepresentable in the compiled matrix, and
inventing a sentinel “any object” id would put wildcard matching on the PEP
decision path (a standing fail-open hazard). So wildcards exist only on
the grant side (Capability::authorizes does directional covering) and
this map expands them at compile time over a closed registry:
mac://<class>/*expands to one rule per registered object of that class — never beyond the registry.- a
*ability expands over exactlySCHEMA_SCOPE_ACTIONS(the generated, closed schema enum). - anything unrecognized (class, name, verb) maps to nothing (fail-closed; never a guessed rule).
Because the expansion is injective — each concrete (class, name, verb)
maps to its own distinct rule, and no two distinct accesses share a rule —
granted_access is the exact inverse of a rule,
which is trivially its most-permissive access: the tightest upper bound of
a singleton is itself. There is no join/LUB logic that could be “too
narrow” and mask an escalation (the #676 risk). The property tests below
pin exactly this: every request that projects onto a rule is authorized by
that rule’s granted_access.
§Registry determinism
Object ids are assigned positionally over the sorted (class, name)
registry, so two nodes constructing the map from the same object list agree
on every id — the compiled matrix and the PEP intern identically with no
side table. The registry is closed at construction: an object added later
has no rule and is denied by default; the remedy is a re-compile at a
new policy generation, not a silent wildcard.
Structs§
- Scope
Permission Map - The production S3-scope permission map. See the module docs for the injective/exact contract and the registry-determinism rule.