v1.7: Enforcement Pyramid + Cost Reduction

Three enforcement gaps identified from the VenLink audit (10 code quality findings) and the NC System retrospective (260K LOC integration gaps), plus a significant cost reduction in the hook pipeline.

Three enforcement tiers shipped

1. Code Quality Enforcement Pipeline

Pushes rule enforcement down the pyramid — from "agent reads the doc" to ruff + AST hooks.

• Reference ruff config with 17 rule sets at `standards/code/ruff-reference.toml`
• Ruff audit mapping 34 standard rules to equivalents (20 direct, 14 guidance-only)
• 2 custom AST hooks: global mutable state (CQ-001) + no-op functions (CQ-002)
• Enforcement annotations on 7 existing standards docs (33 rules annotated)
• 3 new standards: import-discipline, test-isolation, fixture-fidelity
• Compiler `--audit-enforcement` flag validates annotation-to-ruff cross-references

2. Cross-Cutting Concept Registry

Extends the invariants framework with cross-boundary semantic contracts.

• `CONCEPT-NNN` entries in `INVARIANTS.md` with DbC fields (precondition / postcondition / invariant)
• Vocabulary registries for synonym drift detection across bounded contexts
• `check-concepts.sh` verify-phase hook + `check-vocabulary.py` scanner
• 3 VenLink-derived example concepts (tenant ownership, vendor status vocabulary, event handler lifecycle)

3. Integration Seam Test Evidence Gate

Answers the question "for every declared integration boundary, does a test exist that exercises it?"

• `SEAMS.md` manifest for cross-boundary integration seams
• L1 / L2 / L3 evidence levels (exists / exercises both sides / real deps, no mocks)
• `check-seam-evidence.sh` verify-phase hook
• Progressive adoption (no `SEAMS.md` = warning only)

Performance and cost

• Stop hook replaced: Sonnet 4.5 agent → bash `ci-gate.sh`. Zero model invocations on conversational turns. When code is modified, runs pytest / mypy / ruff / invariants mechanically.
• Default gate model: haiku (cheap/fast for mechanical checklists). Sonnet reserved for adversarial review where reasoning matters.
• No pinned model IDs — short aliases (`sonnet`, `opus`, `haiku`) resolve to the latest at runtime.

Other

• `spec-enforcer` registered as agent #21, wired into `/build` Step 7 for adversarial AC verification
• `tasks.py` path fixed for consumer projects (`~/.claude/scripts/tasks.py`)

Numbers

Metric	v1.6	v1.7
Gates	14	17
Hooks	11	14
Agents	20	21
Standards	33	38
Tests	302	409

Upgrade

```
git pull
python3 compile-sdlc.py spec/etc_sdlc.yaml
./install.sh
```