feat(rule_format): add Agent Threat Rules (ATR) format adapter

[eeee2345]

Implements the RuleType abstract contract for ATR (Agent Threat Rules), per discussion with @ecrou-exact in #49.

Files

• app/features/rule/rule_format/available_format/atr_format.py (405 lines) — ATRRule class
• tests/rules/test_atr_format.py (353 lines) — 24 unit tests, all passing

Diff: 758 insertions, 0 modifications to existing files.

Method coverage

All four required RuleType abstract methods + the detect() helper:

Method	Status
format property	"atr"
get_class()	"ATRRule"
validate(content)	Two-layer: YAML parse + ATR semantic (canonical id pattern, 10-category enum, 4-severity enum, agent_source.type enum, detection block shape)
parse_metadata(content, info)	rulezet-canonical shape; preserves ATR-YYYY-NNNNN in original_uuid
get_rule_files(file)	.yaml / .yml extension gate
extract_rules_from_file(filepath)	Single-rule (verbatim, no re-dump), list-of-rules, multi-document YAML
find_rule_in_repo(repo_dir, rule_id)	Match by stable original_uuid or title
detect(content)	Disambiguation helper for the format loader — triggers on ATR-specific markers

The detect() method is implemented as a regular instance method (not @abstractmethod), so the existing RuleType contract in rule_type_abstract.py is unchanged. If/when the abstract grows a detect() method per the discussion in #49, this implementation already satisfies it.

Style alignment with sigma_format.py

• Section header comment blocks (# VALIDATION #, # META PARSING #, etc.)
• Module imports in the same order (stdlib → third-party → app)
• Error-path return shape from parse_metadata mirrors sigma's "Metadata Error" fallback
• detect_cve usage matches sigma's contract: cve_id field is a JSON-encoded string, not a list, so downstream consumers don't need to branch on format
• Single-rule files returned verbatim in extract_rules_from_file (no yaml.safe_dump) — preserves quoting / key-ordering like sigma does

ATR semantic validation

Beyond YAML parsing, validate() enforces:

• id matches ^ATR-\d{4}-\d{5}$
• severity ∈ {critical, high, medium, low}
• tags.category ∈ the canonical 10-category ATR taxonomy
• agent_source.type recognised (additive — unknown types are warnings, not errors)
• detection.conditions is a non-empty array, each entry has field / operator / value
• detection.condition ∈ {any, all, or, and}

Output: ValidationResult(ok, errors, warnings, normalized_content). normalized_content always returns the original input verbatim — never re-dumped.

Tests

$ pytest tests/rules/test_atr_format.py -v
============================== 24 passed in 0.39s ==============================

Covers: format / class identifier, detect() (3 positive + 3 negative cases), validate() (canonical rule accepted + 7 distinct rejection paths), parse_metadata() (explicit CVE list, description fallback, tags flattening, error path), get_rule_files(), extract_rules_from_file() (single-rule verbatim, non-ATR-YAML returns empty, multi-doc).

Tests run inside the existing rulezet test environment (full requirements.txt + Flask app conftest). No additional dependencies introduced.

Auto-registration

Per load_all_rule_formats() in rule_type_abstract.py, dropping the adapter into available_format/ is sufficient — no manual registry edit needed.

On the future rulezet-cast migration

@ecrou-exact mentioned in #49 that the format system is being lifted out to a separate repo (rulezet/rulezet-cast). This adapter is structured to migrate cleanly: zero modifications to existing rulezet-core files, full self-containment of the format logic, and no implicit dependencies on rulezet's internal data model beyond the documented RuleType contract.

Upstream

• ATR repo: https://github.com/Agent-Threat-Rule/agent-threat-rules (MIT, v2.1.2 / 338 rules across 10 attack categories)
• MISP integration: taxonomies#323 + galaxy#1207, both merged 2026-05-10 by @adulau as MISP project lead
• Production deployments: Cisco AI Defense skill-scanner, Microsoft Agent Governance Toolkit (with weekly auto-sync)

Closes #49.