feat(compliance): the EU AI Act and ISO/IEC 42001, mapped clause by clause

[ExposureGuard]

Stacked on #77 (the register adds the agent_register section this maps). Merge that first, then retarget to main.

The evidence pack has mapped to SOC 2 since it existed. That is one framework, and the buyers this product is for get asked about more than one. This adds the two that matter, mapped at the clause level with four rules enforced by tests rather than intentions:

1. "Contributes to", never "satisfies". Every clause names what the evidence does and what it does not cover. A clause with a contribution and no gap reads as a claim that the evidence closes it — test_every_regulatory_clause_states_what_it_does_not_cover fails the build if one is missing.
2. Who the duty falls on is recorded. Article 12 is largely the provider's obligation (build the system so it can log); Article 26 is the deployer's (keep the logs, monitor the system). Haldir's customer is usually the deployer. Collapsing the two tells someone they have met an obligation belonging to whoever built the model — there's a test for that split.
3. High-risk is a precondition, stated as one. Arts 12/26 attach to systems classified high-risk under Annex III. Haldir cannot classify a system, and the mapping says so instead of implying otherwise.
4. Only clauses that can be cited are cited. A.6.2.x is miscatalogued by more than one published catalogue; the mapping uses the numbers confirmed across sources and leaves the rest out. A wrong control number in an evidence pack is worse than a missing one.

Shape

haldir_frameworks.py	The mapping — EU AI Act (Art. 12(1), 12(2)–(3), 26(5), 26(6)'s six-month floor, 14, 15) and ISO/IEC 42001 (A.6.2.8 event logs, A.6.2.6 operation & monitoring, A.6.2.4 V&V, A.9.4 intended use, A.9.2–A.9.3 responsible use)
framework_report()	Renders it for the pack, and groups the readiness score's existing seven checks by the clauses they speak to — no second scoring engine, no second set of signals
The pack	Gains a frameworks section — inside the signed digest, and static, so it cannot drift on its own
Both rendered forms	The mappings render in markdown and HTML, so the auditor-readable document carries the gaps next to the contributions (signature moves to 10)

Unmeasured clauses are reported, not scored. Most of the Act and most of Annex A is organisational; a clause nothing measures returns measured: false with no state. A number invented for it would be exactly the overclaim this module exists to prevent.

Verification

• 1116 tests, flake8 clean, mypy clean over 31 files
• tests/test_frameworks.py also checks the direction that bit while writing it: every section a clause references is a real pack section (it caught proxy and alerting — the latter is the score key for the webhooks section, which is exactly how that slip happens)
• Both renderers inspected; digest verified stable with the new section inside it

---

Sources for the clause citations: EU AI Act Art. 12 logging mandate, Art. 26(6) deployer retention, ISO 42001 Annex A control catalogue.

🤖 Generated with Claude Code