Governed actions: justify or expose the fixed ceilings

Found in the 0.28.0 pre-release review (2026-09-08). Line references are as of main after #926.

Several limits in governed actions are asserted without a matrix row naming the threat, and none is operator-tunable:
- Two Evidence capabilities per action: `MAX_EVIDENCE_CAPABILITIES` in `action_evidence_contracts.rs:17`, also baked into the DDL as `CHECK (ordinal >= 0 AND ordinal < 2)` in `mutation.rs:351`, so raising the constant is a migration. Derive the constraint from the constant or drop the DDL bound and keep the compiler rule.
- Eight concurrent evaluations process-wide: `Semaphore::new(MAXIMUM_EVIDENCE_EVALUATIONS)` in `action_evidence.rs:45`, absent from `EvidenceProviderConfig`, and no test asserts the shipped value. Expose per provider or per runtime.
- The 4096-byte pattern cap in `compiler.rs:1871` is a bare literal written twice with no named constant and no boundary test. Promote it next to the other Rhai and storage limits and test 4096/4097.
- `products/breg/immediate-actions.md:651-660` calls these numbers "Effective defaults" although none is overridable. Rename to fixed limits or make them defaults in fact.

Write slots, refusals, patterns, membership and atomic erasure are already justified by BREG-SEC-36 to 39 and 48; this ticket covers the rest.

Start with action_evidence_contracts.rs:17, mutation.rs:351, action_evidence.rs:45, and compiler.rs:1871, then review products/breg/immediate-actions.md:651-660. Check existing limit and compiler tests before deciding how the named ceilings should be represented. Done means every listed limit is justified or consistently configurable, the 4096/4097 boundary is covered, and the documentation matches the behavior.

Written by the indexing model from the issue text.