Replace #63 stopgap in posthog_feature_attack (tighten TP shape) + consumers must not block trusted first-party skills untriaged

Summary

Two separate problems surfaced when prompt_injection_posthog_feature_attack blocked a first-party skill install. #63 (shipped in 0.2.4) was a stopgap aimed at one phrasing; it does not resolve either root cause. This issue tracks both.

Problem 1 — the rule over-fires on non-instruction framing (rule-level)

The rule matches a verb + feature substring (disable autocapture) anywhere, regardless of the surrounding grammar. That catches informational and even anti-injection prose. Two real first-party shapes trip it:

1. Permissive-modal config docs — "You can disable autocapture entirely by setting autocapture to false." (#63's $benign_config_doc guard suppresses this one.)
2. Negated / prohibitive guidance — "Do not disable autocapture, do not disable session recording." This is guidance to keep the feature on, but the rule matches the disable autocapture substring. #63 does not cover this — it still fires.

#63's guard is also a whole-buffer negative match (any of ($attack_*) and not $benign_config_doc): because a bare string in the condition is true if it matches anywhere, any content containing the benign phrase suppresses the entire rule. A real injection can append a benign-shaped clause to evade. It also pushes consumer-context judgment into the engine.

Proposed direction (rule)

Replace the negative guard with a tightened true-positive shape: require actual instruction framing so neither permissive-modal ("you can disable X") nor negated ("do not disable X") prose matches, while bare imperatives ("Disable X.") still do.

Constraint: yara-x's regex engine (Rust-based) is not expected to support lookbehind/lookahead — confirm during implementation. If so, "match disable X unless preceded by you can / do not" can't be expressed directly; use an imperative/clause-boundary lead-in instead.

Problem 2 — a consumer can hard-block trusted first-party skills (the bigger half)

Warlock only detects; consumers decide what a match means. A consumer's skill-install scan currently terminates on a match without the false-positive triage pass that other scan paths use, so any benign match on trusted first-party skill content blocks the install. Tightening the rule (Problem 1) reduces the noise but does not fix this — the next benign match on any skill re-triggers it. The guarantee that first-party skills don't get blocked lives in the consumer wiring, not the rule.

The consumer-side fix (ensure the install path triages, and/or treat genuinely first-party skills by provenance) is tracked and PR'd in the consumer repo. Design constraint carried there: a trusted-source bypass must key off genuine first-party provenance, not a loose path pattern, or a real injection planted at that path would go unscanned.

Acceptance criteria

Rule (this repo):

• The not $benign_config_doc guard from #63 is removed.
• Neither permissive-modal nor negated/prohibitive config prose matches; bare imperatives and existing positive cases still match.
• No single string can suppress the whole rule across the buffer.
• The false-positive shapes above are locked in as negative-match tests.

Consumer (tracked in the consumer repo, noted here for context):

• A benign match on trusted first-party skill content does not hard-block/terminate the install without triage.
• Any trusted-source bypass keys off first-party provenance, not a loose path pattern.

References

• PR #63 — the 0.2.4 stopgap being replaced.

Start by reviewing PR #63 and the prompt_injection_posthog_feature_attack rule, then inspect its existing positive cases and the rule-level tests. Confirm the yara-x regex constraints before choosing the instruction-framing approach. Done means removing the benign-config guard, preventing permissive-modal and negated prose matches, preserving bare imperatives, and adding negative-match coverage; consumer wiring is tracked in another repository.