v3.21.0

Added

One judgment-only rule. Catalog goes from 60 to 61 detection categories; the engine stays at 46 types.

• Narrated candor — announcing your own disclosure instead of disclosing: "Two caveats I would rather flag than let you discover later:", "I want to be upfront:", "rather than bury this", "I could have left this out, but". The content is "Two caveats:"; the rest advertises the writer's forthrightness. Completes a set with two existing rules: chatbot artifacts perform helpfulness and sycophantic tone flatters the reader, while this performs candor about oneself. Usually arrives as a matched antithesis (flag rather than let you discover), which is a tell in its own right. Added to the P1 severity tier.
• Implemented as a detector, then reverted before release, following the precedent set by wall-of-text replies. An adversarial review of the first implementation found it flagged idiomatic conflict-of-interest disclosure ("In the interest of full disclosure, I own shares in the company discussed in this article") and the ordinary English comparative ("I'd rather fix it than let you inherit the mess"), and that the bounded-repeat pattern backtracked catastrophically on long \w runs — 36 seconds on a 4 KB input. Every regex tight enough to spare the carve-outs stopped matching the tell, so the rule ships as skill prose only. detector/CATEGORIES.md §C records the reasoning.
• Two carve-outs are now explicit in the rule, because they are the cases the failed detector proved are hard: conflict-of-interest disclosure keeps its conventional opening, and the ordinary comparative is not this pattern. The tell requires that what follows the frame is the disclosure itself.