feat(library): add Agent Threat Rules (ATR) detection rail

[eeee2345]

Description

Adds an input rail that matches the user message against Agent Threat Rules, an open MIT-licensed detection catalog for AI-agent attacks (prompt injection, jailbreak, tool poisoning, MCP attacks, skill compromise). Rules ship inside the optional pyatr package and are evaluated in-process, so the rail needs no API key, no service endpoint and no network access, and the same input yields the same verdict on every run. It is modelled on injection_detection, the other local optional-dependency rail.

This supersedes #1992, which predates the rail-manifest migration and could not have merged in its shape — a library package with actions and no rail.py fails test_library_action_packages_declare_manifests. Rebasing was not enough; the module needed rewriting to the manifest pattern.

Areas worth careful review:

• rail.py declares RailPrivacy() with every disclosure flag false and no remote services. Please sanity-check that this is the intended way to express a rail that contacts nothing, rather than simply leaving the field at its default.
• actions.py returns RailOutcome and carries matched rule IDs plus max severity in metadata on allow as well as block, so sub-threshold matches remain visible in traces. Say the word if you would rather allow paths carried empty metadata.
• schemas/rails_config.snapshot.json is regenerated. The diff is purely additive (+25 / -0) and contains only the atr config key and ATRDetection definition.

One defect found while porting, flagged separately because it is the kind of thing that would have been unpleasant in a release: the Colang 1 flow in #1992 was declared define flow rather than define subflow. As a top-level flow the runtime activated it independently of configuration — it ran for configs that never enabled the rail and overrode the configured rail's refusal message, rendering an ATR message with an empty rule list. Locally that showed up as injection_detection_reject_blocks_injection failing with an ATR refusal string in place of the injection one. Both flows now emit the shared bot refuse to respond intent instead of an inline templated bot say, and test_colang_1_flow_is_declared_as_a_subflow guards the declaration.

block_severities: [] is honored as monitor-only; only an absent config section falls back to ["critical", "high"].

Related Issue(s)

• Add an Agent Threat Rules (ATR) detection library rail #1991 — I have asked there for triage and assignment, since the issue currently has no labels or assignee. Happy for the answer to be "not a fit"; that closes it cleanly and I will withdraw this PR.

Verification

Run locally against develop @ origin/develop, Python 3.12, pyatr 0.2.7:

• tests/test_atr_rail.py, tests/manifests/, tests/rails/llm/, tests/test_runtime_flow_gate_equivalence.py — 399 passed.
• Baselined every failure against clean develop before treating it as mine; all four initial failures were caused by this change and are fixed, not suppressed.
• ruff format --check and ruff check clean on the touched files.
• uv lock --check passes.

Not run: the recorded/live-credential suites and the docs build.

Residual risk: uv.lock is regenerated with uv 0.10.1. The marker normalization and exclude-newer churn in that file comes from the uv version rather than from this change — uv-latest.yml already runs uv lock --check on latest uv, but if you pin a different version please regenerate.

AI Assistance

• No AI tools were used.
• AI tools were used; a human reviewed and can explain every change (tool: Claude Code).

Checklist

• I've read the CONTRIBUTING guidelines.
• This PR links to a triaged issue assigned to me. — Add an Agent Threat Rules (ATR) detection library rail #1991 is open but not yet triaged or assigned; requested in the issue.
• My PR title follows the project commit convention.
• I've updated the documentation if applicable.
• I've added tests if applicable.
• I've noted any verification beyond CI and any checks I couldn't run.
• I did not update generated changelog files manually.
• I addressed all CodeRabbit, Greptile, and other review comments, or replied with why no change is needed. — none yet on this PR.
• @mentions of the person or team responsible for reviewing proposed changes. — @Pouyanpi

[eeee2345]

@tgasser-nv — this has been sitting in needs triage for about two weeks, so a quick pointer to help it find a reviewer.

It adds one local input rail under nemoguardrails/library/atr/, modelled on injection_detection. Optional pyatr dependency, evaluated in-process, no API key, no endpoint, no network, deterministic verdicts. The blast radius is that new folder plus one optional-dependency line in pyproject.toml; nothing in the existing rails changes. CI is green and it is currently mergeable.

If a smaller diff would be easier to review, I can pull the docs out into a separate PR, or rebase if it has drifted. Whatever lowers the review cost.

[github-actions]

PR merge guidance

@eeee2345 thanks for the PR. GitHub is currently blocking merge for one or more repository requirements:

• 1 commit does not have a verified signature (c1ad48c). Please sign the commits and force-push the updated branch.

Relevant guide:

• Signed commits: https://github.com/NVIDIA-NeMo/Guardrails/blob/develop/CONTRIBUTING.md#commit-signing
• Contribution guide: https://github.com/NVIDIA-NeMo/Guardrails/blob/develop/CONTRIBUTING.md