feat(manifest): add per-file digest provenance - #87

Summary

Closes #84.

• Add input_file_digests and config_file_digests to every primary run manifest.
• Hash each file's exact shipped bytes with SHA-256; no YAML parsing or reserialization.
• Normalize keys to repository-relative POSIX paths and emit them in lexical order.
• Preserve the existing input_digest and config_digest aggregate contract.
• Pin manifest-covered shipped text inputs/configs to LF with .gitattributes so exact-byte values reproduce across Windows and Linux clean clones.
• Keep structured artifact diff reporting (Add structured artifact-contract diff reporting #85) out of scope.

Design decision

The legacy aggregate digests remain label-sorted and LF/CRLF-normalized for UTF-8 text. Per-file digests use read_bytes() directly and are keyed by canonical repository-relative paths. A bounded LF checkout contract makes those shipped bytes platform-stable; the per-file algorithm itself does not normalize line endings. The schema keeps the new maps optional so older run-manifest/v1 consumers remain readable while current writers always emit them.

Main risk

The LF checkout contract renormalizes four historical CRLF/mixed tracked input/config files at the Git blob boundary. Parsed content is unchanged, aggregate digest values remain unchanged, and the resulting per-file values are now reproducible across clean clones.

Compatibility impact

Additive manifest fields only. Existing aggregate digest values are asserted against the v1.2 contract, a legacy manifest without per-file maps remains schema-valid, and current writers emit exact-byte maps for all six primary manifests.

Rollback path

Revert the four commits in order, or stop consuming the optional per-file maps; input_digest and config_digest remain unchanged. If the LF checkout contract must be reverted, regenerate manifests on the chosen repository byte contract before publishing.

Validation

• python -m pytest -q — 196 passed locally.
• PR CI — 2 test jobs and 2 Repo Sentinel jobs passed.
• python scripts/regenerate_artifacts.py --check — 29 strict artifacts matched; 6 visual snapshots smoke-generated.
• Independent clean-clone consumer reproduction — all 6 manifests rehashed from read_bytes() with sorted path checks passed.
• Exact-byte, line-ending sensitivity, lexical ordering, multi-file sensitivity, and legacy compatibility tests pass.