Make SI check intent before work and keep authorized work moving

[infotradescout]

Selective Intelligence could begin Worker work before checking the user's intended outcome, while a refusal for one tool action could reopen authorization for unrelated work. This PR makes Orchestrator, Worker/Builder, and Objector standing responsibilities: reconstruct intent, challenge a plausible material misreading before Worker dispatch or work, execute the bounded task, and review the result and proof before reporting. A denied action stays scoped to that action; clear reversible work continues. Formal Council roles remain conditional.

The canonical skill, ChatGPT adapter, public plugin description, fallback guide, discovery files, and README now carry the same role order. Checkpoint saving uses short atomic temporary names so deep Windows worktrees can save and resume. The public runtime keeps its 61-entry limit and required activation and execution files; two optional references remain in the full source release but are omitted from that runtime bundle.

Validation on clean native Linux source 6b9cc1e9fa13c31398e6f874ec4a42fe97acaf04: the unchanged SI quality gate passed all five checks. The parent runtime commit 5fbdbc42d86ad6a5a4520ec47369237f3944e304 also passed discovery bridge 11/11, native pointers 8/8, ChatGPT adapter 59 files with one skill entrypoint, public plugin 4/4, and extracted adapter delivery 25/25; the follow-up commit changes only README. Independent Objector reviews passed the role contract, runtime integration, and README correction. GitHub's Quality job for 6b9cc1e failed before any job steps ran, so hosted execution is absent. Fresh model-driven behavior and all-client acceptance remain separate; this PR stays draft and unmerged.