Evidence and hierarchy (issue #19)

[mxriverlynn]

Summary

This PR extracts the evidence-based reasoning rule out of /research and into a canonical reference that every judgment-producing skill and reviewing agent loads at runtime, so the trust-class vocabulary and corroboration gate apply uniformly across the plugin instead of being one skill's private convention.

• Adds a third foundational mechanic to the plugin alongside sizing and YAGNI: an explicit posture on what counts as evidence (codebase / web / provided trust classes), how to characterize its strength (proximity heuristic, corroboration gate scoped to web sources), and what to do when no evidence exists at all (label and defer with a reopen trigger, same pattern YAGNI uses for deferred items).
• Scope is documentation-only — no runtime code changes, no SKILL frontmatter or allowed-tools shifts. Every changed file is markdown: one new canonical rule under plugin/references/, one new operator-facing long-form doc, durable research backing the rule, and a sweep of seven SKILL.md files plus four long-form agent docs adding a single-line pointer to the rule.
• The interesting design call is the asymmetry: the corroboration gate fires on web sources but not on codebase findings. A single path:line citation stands; a single Stack Overflow answer driving a recommendation does not. Worth weighing whether that asymmetry holds up at the seams (codebase-vs-web disagreement, provided-material handling).
• Reviewer attention is best spent on the rule's wording and the boundary with YAGNI, not on the per-skill pointer edits, which are mechanical.

What to look at first

• The asymmetry between web and codebase corroboration in plugin/references/evidence-rule.md — Principle 2 explicitly scopes the gate to web sources and defers extending it to codebase evidence. The rationale is that a single file path is the current state of the system, but it is worth pushing on whether that holds when the codebase finding is itself a single grep hit in unfamiliar code.
• The YAGNI-vs-evidence boundary in evidence-rule.md ("When the rule and YAGNI both apply") and the same section in docs/evidence.md. The claim is that YAGNI gates inclusion and this rule characterizes quality once included, and that the two do not collapse. Worth checking the seam holds at edge cases like a YAGNI-passing item whose only evidence is a single uncorroborated blog post.
• The proximity-inversion conditions (formal methods, regulatory, pre-incident observation of intended behavior). These are the carve-outs that prevent the principle from being misread as "running code always beats docs." Confirm the three named contexts cover the cases that come up in /research and /investigate today.
• The 728-line research report under docs/research/evidence-hierarchy.md is the durable backing for the rule but is not on the critical path for review; a skim of its conclusions is sufficient unless something in the rule looks wrong.

Files of interest

• plugin/references/evidence-rule.md — the new canonical rule consumed at runtime by every evidence-aware skill and agent; this is the file that matters most.
• docs/evidence.md — the operator-facing long-form companion; verify it stays consistent with the canonical rule and follows the writing-voice profile.
• docs/research/evidence-hierarchy.md — the research that justifies the principles, the trust classes, and the scoped corroboration gate.
• plugin/skills/research/SKILL.md — the representative skill update; /research is where the vocabulary originated, so its diff shows the pattern other skills follow.
• docs/plans/canonical-evidence-resource/feature-implementation-plan.md — the plan that drove the change, useful if a reviewer wants to check scope decisions against intent.