feat(evidence): scope the configuration revision per requirement

[jeremi]

Why

configurationRevision was one SHA-256 over every file in the acceptance bundle. A relying party pins it, verification compares it by exact equality, and every policy mismatch collapses to one generic VerificationError::Policy. So any byte change anywhere in the bundle invalidated every relying party's pinned value at once, and each of them saw an opaque failure until they had re-reviewed and redeployed.

What changes

An assertion now carries a digest over the configuration and artifact closure of the one requirement it answers.

• Bundle::revision() stays the deployment's own identity: audit bundle_revision, /status, evidence check, operator diagnostics.
• Bundle::configuration_revision(requirement_id) is what an assertion carries. Assertion construction, discovery, local verification, and the fixture CLI path all use it. Every call site now declares which of the two it means.
• Discovery publishes it per EvidenceDefinition instead of once per document, so a relying party pins the requirements it consumes rather than the deployment.

A requirement's closure is its own configuration plus its source, selector profiles, authority grants, and the scripts, schemas, codelists, and fixtures those reach. Retired public verification keys stay in every closure, so key-rollover coverage is unchanged.

The canonical projection starts from serde_json::to_value(&EvidenceConfig) and narrows only the four per-requirement members, so a configuration member added later is covered by default rather than silently omitted. Canonicalization is the shared RFC 8785 implementation (registry_platform_crypto::canonicalize_json), not a bespoke serializer.

Security review notes

• Assertion construction and verification policy comparison change value, not shape. What the signature covers is unchanged.
• Tripwire narrowing is intended and is the point of the change: an edit outside a requirement's closure no longer invalidates its pinned revision. Retired public verification keys stay in every closure, so key rollover still moves every revision.
• The projection is faithful because the configuration types deny unknown fields, so no authored member can sit outside the serialized value it starts from.
• The digest is domain separated (registry.evidence.requirement-revision/v1\0) and length prefixed per path and per byte string. The projection rides in the closure under evidence.yaml#requirement, and artifact paths cannot contain #, so it cannot collide with a real file.
• Every currently pinned revision changes once. Relying parties must re-pin from authenticated discovery.
• Moving the discovery field from the document to each definition is a breaking wire change. Both client SDK types deny unknown fields, so a deployment still publishing it at the document level is refused rather than silently ignored, and a relying party cannot pin a value no assertion carries.

Area boundary

This spans three owning areas on purpose: the frozen-contract amendment (products/), its implementation (crates/), and the reader-facing prose and commands (docs/site/). The application tutorial's pinning script read published["configurationRevision"] at the document level, so leaving it would have documented a procedure that no longer verifies.

Verification

• cargo fmt --check, cargo clippy --offline --workspace --all-targets -- -D warnings: clean.
• cargo test --offline --workspace: 188 test result: ok, zero failures.
• products/evidence/scripts/check-contracts.sh, check-source-neutrality.sh, check-verifier-portability.sh: all pass. Generated artifacts were regenerated by their documented generator, not hand-edited.
• docs/site: npm test (316 pass) and npm run check (full gate, including the evidence-tutorial drift check, 35973 links).
• Node binding: npm ci, npm run build:debug, npm test (22 pass), npm run check:types, cmp LICENSE.
• Python binding: extension-module build, python3 -m unittest discover -s tests/python (33 pass), cmp LICENSE.
• docs/site/src/content/docs/tutorials/request-evidence-from-an-application.mdx is not in the tutorial replay gate's registered set. Its pinning script was replayed against its own documented discovery sample and reproduces its documented procedure.json byte for byte.
• The two #[ignore]d live gates in crates/registry-evidencectl/tests/production_handoff.rs now read the expected revision from authenticated discovery, so the assertion check stays independent of the signed payload it is compared against, and assert it differs from the bundle revision.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c15bbb7b1c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".