Add graduated multi-agent research integrity

[Silentpartnercoding]

What changed

• add concise root instructions for agents and contributors
• make Claude import the canonical AGENTS.md instructions instead of duplicating them
• enforce that bridge with regression tests so agent-specific instructions cannot silently drift
• define routine, exploratory, candidate, canonical, imported, and authority-sensitive lanes
• add a one-page contributor quickstart with a concrete example for every lane
• add one make verify command shared by local contributors and CI
• add safe record scaffolding for exploratory, candidate, canonical, and imported work
• add actionable Fix guidance without automatically rewriting evidence
• add per-record lifecycle metadata and a diff-aware integrity checker
• add planted-defect tests for promotion, independence, immutability, authority, instruction alignment, and scaffolding misuse
• clarify that clean reproduction is not independent verification
• add an advisory PR template and ownership routing

Contribution boundary

Routine code, documentation, adapters, and exploratory research remain lightweight. Stronger requirements apply only when work is promoted into candidate or canonical research. Existing untouched research is grandfathered.

Multiple agents may collaborate, but agents under one controller remain one control domain. Verification cannot create independence or authority that the evidence does not already establish.

The record command computes hashes, resolves commits, refuses unsafe overwrites, and requires a candidate record to be committed before its canonical result. It does not silently repair or promote evidence.

Validation

• make verify passed
• 301 Python tests passed; 8 skipped
• site lint, build, and render test passed
• research-integrity check passed
• public-boundary and withheld-material checks passed
• 4 preregistration bindings verified