feat: add Compound Marketing operating loop

[douglas951]

Summary

Marketing OS can now carry one governed decision record from company context through
strategy, GTM, human review, and durable learning. The system preserves human rulings
across handoffs and refuses to call a workflow "compounding" until a comparable real run
meets the declared proof threshold.

This PR builds on the Every master-brand canon in #19. It establishes the core
strategy, marketing, and GTM layer; design and channel automation remain outside V1.

Design decisions

• Every project does two jobs: produce effective marketing now and improve the system
  that produces the next project.
• The governing loop is context -> strategy -> market -> memory -> stronger next run.
• One shared run directory and decision record keep source authority, protected language,
  open questions, and approved decisions intact across specialist skills and sessions.
• Every applicable stage produces a human-signed artifact receipt. The proof checker
  resolves those shared files and rejects missing receipts, artifact drift, decision-ID
  drift, record-version drift, and paths that escape the run directory.
• Agent pods are optional infrastructure. Agent count and autonomy labels are not proof.
• Structural, operational, and compounding proof are separate. Only the comparison
  command can return proof_ready: true.

The Brand Book and Thesis GTM cases are real retrospective baselines. Both intentionally
remain below the compounding-proof threshold.

Validation

• 40 repository tests pass, including 27 Compound Marketing validator and CLI tests.
• The Marketing OS learning-loop smoke suite passes.
• Python compilation, JSON/YAML parsing, path checks, and diff checks pass.
• Two fresh-context agent runs preserved the strategy and GTM decision contracts.
• Correctness, testing, maintainability, agent-native, efficiency, reuse, and adversarial
  review findings were fixed. The independent Claude route could not authenticate, so
  the review used the required in-process adversarial fallback.

New concepts

Compound Marketing

Compound Marketing treats the reasoning behind good marketing as reusable operating
capital. A run records the decisions that governed the work, measures the human burden
and defects, and promotes only the learning that survives real use. It should not be
used for trivial edits or as a claim about agent autonomy.

Artifact receipt

An artifact receipt ties an accepted stage artifact to its run, decision-record version,
preserved decision IDs, named verifier, and verification time. A publishable proof claim
fails when any required receipt is missing, unreadable, or inconsistent.

Post-Deploy Monitoring & Validation

• Owner: Douglas Brundage
• Window: The next comparable strategy or GTM run, followed by one independent
  teammate rerun within 30 days
• Watch: Human review minutes, first-pass acceptance, critical defects, accepted
  inherited decision IDs, artifact receipts, and duplicate run creation
• Healthy: The follow-up run inherits decisions, requires at least 20% less
  correction, and shows no acceptance or defect regression
• Failure: An incomplete record qualifies, a receipt cannot be resolved, a fresh
  session cannot resume the run, or decisions and protected language drift between artifacts
• Mitigation: Stop the claim, retain the records as a baseline, fix the governing
  contract or validator, and continue with the existing specialist skill

Related: #19

---