[ravenoak]

The generator fails on every image, the card API and the training tool the project uses are gone or going, and the trainer, label catalog, subset selection, evaluation and inference are neither specified nor built. DESIGN.md covered only the generator.

This adds the documents that define the work, and the backlog that does it. The package code does not change. DESIGN.md is removed, and README.md, CLAUDE.md and AGENTS.md point to the new docs.

Documents (docs/)

• intent.md replaces DESIGN.md: purpose, users, outcomes, proposed success measures, non-goals, constraints, glossary.
• principles.md, architecture.md.
• specs/: catalog, generation, selection, training, evaluation, inference, and cross-cutting. 83 EARS requirements with FR-/NFR- ids, each with a check. Each spec ends with its current state and the issue that closes each gap.
• adr/: 0001 to 0004 Accepted (record decisions, PyTorch + timm, retrieval first, TCGdex source; 0002 to 0004 follow your answers to the planning questions). 0005 (plain-file formats) and 0006 (offline plus online augmentation) are Proposed.
• rtm.md: every requirement with its issue, implementation and test refs.
• examinations/2026-10-09-critical-examination.md: the 37 finding records, the verification, the commands behind each count, the limits, and the corrections from a second review.

Backlog: issues #75 to #111, in four milestones (M1 Pipeline runs, M2 Catalog and selection, M3 Trainer v1, M4 Evaluation and inference). Each cites its ReqIDs and acceptance criteria. Links from issues to docs/ resolve once this merges.

Needs the owner

• Accept or change ADR 0005 and ADR 0006. Add a label catalog with a versioned schema, replacing pickles #88 (P0) waits on 0005 and Add a manifest-driven dataset class and the train command skeleton #101 on 0006.
• Confirm the success measures in docs/intent.md. They are proposals with no baseline.
• Decide how the real-photo set is collected (Add real-photo evaluation with per-label metrics and reports #106).
• infra-local-ci (prek.toml runs pytest -n auto, which the local-CI standard forbids) is recorded in the examination and not filed. This repository's CLAUDE.md requires -n auto.

Checked

• uv run prek run --all-files passes: ruff, pyright, and pytest (42 passed, 1 skipped).
• A local script (not in the repo) confirmed that all 83 requirements appear in docs/rtm.md, that relative links resolve, and that no unknown ReqID is cited. Add a check that links tests to requirements #99 adds a repo check.
• A second reviewer found defects in the first version. They are fixed and listed in the examination.
• Not verified: TCGdex rate limits, image rights, and the official Standard rotation date.