v3.2.0: conventions gate, config dogfooding, Agent Plugins manifest

Sharpens the red-team gate and hardens the repo's own conventions, validated by a defect-injection eval (11 arms: A/B on planted and clean diffs, plus a self-application arm), two independent release reviews, and a run of the real devloop:red-team against the release diff. The eval also refuted two candidate additions (an external-compat angle and blindspot-probe handling): the shipped gate already caught those defect classes, so neither shipped.

Added

• Landing site under docs/. A GitHub Pages site (index.html, llms.txt, FAQ). Shipped after 3.1.1 and carried in this release.
• devloop dogfoods its own config convention. The README tells users to commit .devloop/config.md and domain.md, but the repo gitignored the whole directory, so red-team's conventions angle returned nothing on every self-review it had ever run. Both are now committed, alongside a thin harness-native AGENTS.md that points at them. config.md's Standards carry only rules validate.sh cannot mechanize, each traceable to the release that fixed it.
• Vendor-neutral Agent Plugins manifest. A root plugin.json in the Agent Plugins 1.0.0 format (backed by Amazon, Cursor, Microsoft, OpenAI, and Vercel; Kiro has announced support), alongside the existing Claude Code and Codex manifests. skills/ was already conformant. Claude Code and Codex read their own manifests and ignore the root one (verified with claude plugin validate --strict); conformance is validated against the published JSON Schema, not a tested install.
• validate.sh sections 19 to 21. Section 19 guards the dogfooded config; section 20 pins the red-team angle counts across the README, docs, SKILL.md, and the mode table; section 21 validates the new manifest and pins version and description across all three manifests. The suite is now 244 checks, up from 219.

Changed

• Convention faults now block the gate. The Phase 4 verdict ladder FAILs on a CONFIRMED conventions finding whose failure scenario is a genuine fault (security, data loss, a broken documented invariant, or a robustness gap such as a missing rate limit or an unbounded query); purely stylistic findings (naming, commit-message format, doc location) stay WARN. Previously only CONFIRMED correctness could FAIL, so a real standards violation with no correctness face could pass the /implement gate. The gate wording and the README/docs match.
• No fixed cap on findings. Removed the numeric per-angle limits in red-team in favor of "surface every candidate with a nameable failure"; the no-pad discipline is unchanged.