task: consolidate shared contracts and packaged-resource access #451

Goal

Consolidate existing shared data/byte contracts and packaged-resource access into explicitly bounded namespaces. This is physical ownership cleanup, not a new contracts framework or a generic common/utils package. It closes root-level scattering and does not include an internal refactor of existing subpackages. Related: #387, #408 and #412.

PR checklist

• C1 — Move shared data/byte contracts and retire old root modules. Put the existing canonical JSON implementation, RunSpec and claim types under inferrsi.contracts as contracts/json_v1.py, contracts/runspec.py and contracts/claim.py; replace contracts.py with that package, retaining from inferrsi.contracts import RunSpec and keeping contracts/__init__.py limited to that one public re-export so there is exactly one RunSpec initialization path. In this same PR delete root _json_v1.py, runspec.py, claim.py and contracts.py; update active consumers, exact-path policy declarations, docs and tests. Add the bounded exact-root inventory guard here; initialize it to this PR's real post-migration file set, then shrink it in every later removal PR (C2, R1, A1/A2, M1), rather than waiting until R1 to prevent new flat modules.
• C2 — Move resource access and separate evidence rules from host I/O. Move the neutral repository-pack facade to resources/repository_packs.py (inferrsi.resources.repository_packs); put evidence segment/path rules in contracts/evidence.py (contracts.evidence) and directory creation in apps/evidence_layout.py. In this same PR delete root _repo_pack_data.py and evidence_paths.py; update consumers, manifest/technical-leaf rules, docs and tests. Keep path construction in contracts.evidence free of filesystem writes; mkdir stays only in apps.evidence_layout, with a focused test proving the split. In docs/legacy-knowledge-manifest.json, relock the technical_leaf entry for the moved file (module, file path, file_sha256, allowed_imports, top_level_shape) and repoint the catalog / infra.inspection compatibility targets from inferrsi._repo_pack_data to inferrsi.resources.repository_packs; per the manifest's resource_path_revision note this relock goes through formal architecture review. Keep the files("inferrsi").joinpath("harness", "data", "repo_packs", ...) anchor unchanged so the leaf still imports no inferrsi module, and keep resources/__init__.py import-free.

Dependencies and boundaries

C2 uses the contracts package established by C1. These two PRs must remain independently testable; do not split a move from its consumer updates. No changes to task evaluation, format retirement, model experiments, or ownership of the K02 data files. Keep package data physically under its existing Knowledge/Harness/Benchmark/Apps owners.

Acceptance

• Canonical JSON bytes and error stages, RunSpec validation/round trips, Verdict identity and Harness re-exports retain their existing semantics.
• Contract-package initialization imports no application or core implementation and performs no filesystem I/O; import inferrsi.contracts is safe and side-effect-free. The repository-pack facade remains a neutral read-only leaf: it must not import Harness to find data.
• Preserve the one REPO_TYPES object and the bytes/access semantics of all seven repository packs; assert object identity across the remaining consumers (for example catalog.REPO_TYPES is resources.repository_packs.REPO_TYPES while the root shell still exists) and verify installed-wheel access outside the checkout.
• Evidence path construction stays separate from mkdir. The retirement note lists old module paths -> new paths and states the exact external pickle/class-module-path impact, rather than adding indefinite aliases.
• Root tracked-file inventory decreases from 27 to 23 after C1 and to 21 after C2, subject to refreshing the implementation base; record exact file lists, not just counts. The build-generated _version.py is not a tracked root business file and is allowed only as wheel metadata.

Delivery discipline

• One implementation Claude session handles only the current PR in an isolated worktree. A second Claude session may only plan the next PR in a separate planning worktree; it must not edit source.
• Move implementation, update all active consumers/docs/tests, and delete the corresponding old root modules in the same PR. Do not introduce permanent compatibility shells or module-registry aliases.
• Preserve CLI command behavior, original evidence bytes, record/format contracts and isolation/security checks. Announce retired Python import/serialization module paths; historical frozen evidence remains pinned to its original commit or wheel, not rewritten.
• Before each implementation or review handoff, refresh main and confirmed dependencies. A completed plan is not implementation or approval. No automatic merge.