feat: add claim and replication schemas — the core thesis objects

[cschanhniem]

Summary

The repo's thesis is "CI/CD for truth claims": make false AI-generated knowledge harder to merge. But the protocol had no persistent claim object — only transient submissions and single-source evidence records. This PR adds the missing core objects.

Changes

New schemas

• schemas/claim.schema.json — the core protocol object. A persistent, falsifiable statement with a verification lifecycle (unverified → dry-lab-verified → needs-replication → replicated → accepted → field-tested), evidence links, failure modes, a kill condition, and required reviewer roles. High-safety claims cannot advance without red-team review. Falsified and deprecated are terminal states; falsified claims stay in the record so contributors do not repeat them.

• schemas/replication.schema.json — independent replication records with replicator identity (must differ from original submitter), environment (runtime, OS, packages), input hash, result enum (confirmed / partially-confirmed / failed / inconclusive), and divergence tracking. Self-replication is not replication.

Strengthened schemas

• schemas/evidence.schema.json — added model-prediction, computational-analysis, and wet-lab-confirmation evidence types. The thesis envisions computational claims (molecule candidates, theorem sketches, benchmark results) but the enum had no way to represent them. Added optional doi and archive_url fields to insulate records against link rot.

• schemas/agent-submission.schema.json — added kill_condition as a required field. SKILL.md pattern Add agent merge bar note #3 already required it in prose; the schema now enforces it. Protocol drift fixed.

New examples

• examples/claim.example.json — worked example using the dengue-heat-vietnam pack
• examples/replication.example.json — worked example with a partially-confirmed replication result
• examples/agent-submission.example.json — updated to include kill_condition

Validator

• scripts/validate-repo.mjs — compiles and validates all new schemas and example files

Docs

• AGENTS.md — replaced stale 21-pack table with pointers to generated indexes (actual count: 100). Added claim lifecycle section.
• ROADMAP.md — fixed "15 problem packs" to "100 problem packs"
• CLAUDE.md — updated schemas table and evidence type list
• schemas/AGENTS.md — updated component and flow diagrams
• examples/AGENTS.md — added new example files to key components
• agents/literature-scout.md — updated evidence types table with new types
• README.md — updated repository map description

AI-slop cleanup

• Replaced vague metaphorical "landscape" in the skills-training pack with precise terms ("full scope of skills-training availability", "training-system assessment")
• Legitimate ecological uses of "landscape" (landscape simplification, agricultural landscapes, mosaic landscapes) in biodiversity and climate packs were left intact

Protocol Notes

• The kill_condition field in agent-submission.schema.json was required by SKILL.md pattern Add agent merge bar note #3 but absent from the schema. This is exactly the kind of prose-rule-vs-schema drift the protocol should catch. A future validator rule could scan agent guides for required fields and check them against the corresponding schema.
• The stale pack count in AGENTS.md (21) and ROADMAP.md (15) vs actual (100) suggests the validator could check that hand-written files do not embed pack counts, or that they match the generated index. Currently the rule exists only in prose ("Never trust an embedded count").

Verification

• pnpm validate — passes
• pnpm reproducibility:check — passes (516 tasks)
• pnpm build — passes (100 packs)
• pnpm format:check — passes