v0.14.0 - runnable recipes

Nine recipes you can actually run. The library has advertised nine skill chains since v0.11.0 - stress-test a decision, expand your options, audit a piece of reasoning - and until now "running" one meant opening a markdown file and executing its skills by hand, two to four of them depending on the recipe, carrying the artifact between them yourself. Each is now a single command. No new frameworks; the catalog stays 56 evidence-graded core skills plus 7 contested lenses.

For everyone

• /think-stress-test-decision, and eight more. Every recipe now runs as one invocation, executing the whole chain with its documented handoff so the compressed artifact moves between steps for you instead of via your clipboard.
• A troubleshooting page, in the words you would actually see. The plugin installed but the command does not appear; the skill fired but gave you prose instead of an artifact; the recommendation felt wrong. There is now a page for that, and every entry opens with the verbatim output, harvested from the tools rather than paraphrased, so you can search for the string in front of you.
• Both quickstarts show what success looks like. A real (truncated) risk register from the premortem, so you can tell a good run from a vague one - which is also what made "I got an essay" hard to self-diagnose.
• The seven contested lenses are reachable by name. They all existed and none was linked from the samples index, so browsing for SWOT dead-ended. A library that only demonstrates its strongest methods is showing a highlight reel.

For builders and contributors

• The recipe commands are generated, never authored. A new conformance-gate layer regenerates all nine and fails on any byte of drift, so a hand-edited command cannot claim to run a chain it no longer matches. The recipe body is carried verbatim, because the numbered handoff lines are the authored value. The gate goes 14 -> 15 layers.
• The library's description of itself is true now. All 67 skill sidecars described a shipped, active skill as an unshipped draft, and described the 63 that carry two eval stamps as unmeasured. (The other four are the meta-skills, which sit outside the behavioral eval run by design and say so.) Both copies of the applicator engine also told the model the catalog held 34 skills when it holds 63. Both are now gate-asserted rather than hand-maintained, so neither can drift back.
• Five reproduced contested-guard bypasses are closed. The caveat-first contract is this library's flagship claim; an adversarial pass found five structural ways around the code enforcing it. All five are closed or named as documented residuals, verified red before the fix, with all seven shipped lenses passing unchanged - the criterion that guards against over-tightening.
• The guards now guard themselves. The three gate layers with no tests of their own gained 38, and every generator reports a tally instead of a bare "OK", so a run that processed zero files stops reading as success. The suite goes 137 to 235.
• Security and supply chain. The Framework Advisor no longer tells the model that pasted text is authoritative - it is evidence to analyse, not instructions to follow. On dependencies, the honest position rather than the flattering one: site/ vulnerabilities were taken from 10 to 6 without a breaking change in August, and as of this release npm audit reports 10 again (1 critical, 4 high) - new advisories against the same pinned versions, including a critical Astro XSS. Nothing in site/ changed; the advisories did. These are build-time dependencies of the docs site, so the exposure is a developer machine and CI rather than anyone who installs the plugin, and the fix gets its own PR rather than riding along inside a release cut.
• The conformance claim names its Standard version, and the numbers behind it are now stated once with their provenance rather than copied onto six pages, where they twice went stale. The pin and the declared Standard are documented as two separate dials, because conflating them is the easy mistake.