[aviggiano]

Summary

• publish typed coverage evidence with distinct Recon-selected and full-production declaration-completeness views
• preserve LCOV and Recon-selection snapshots and verify source, range, attribution, and hit consistency at publication
• disclose exclusions and zero-coverage components, with typed unavailable evidence for blocked measurements
• carry the exact handoff into report@3 and one canonical scoped Markdown section, rejecting contradictory or unscoped coverage claims
• add document-local artifact validation so producers catch cross-field errors before publication

Raw covg-eval results remain iteration-only and do not define either published declaration-completeness view.

Fixes #532

Validation

• pnpm -w run ci

Greptile Summary

The PR adds typed scoped coverage evidence and carries it through artifact validation, runtime authentication, report publication, evals, and public bundles.

• Adds measured and unavailable coverage-evidence contracts with document-local reconciliation.
• Authenticates production declarations, Recon selection, LCOV attribution and hits at runtime.
• Adds canonical Markdown projection and exact report handoff.
• Introduces producer-facing artifact contract validation and updates topology, schemas, prompts, and documentation.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
packages/artifacts/src/coverage-evidence.ts	Defines and reconciles typed measured or unavailable coverage evidence, including file, range, scoped-view, exclusion, and zero-coverage relationships.
packages/artifacts/src/artifact-contracts.ts	Registers the coverage evidence and updated report and coverage-goal contracts with document-local semantic validation.
packages/cli/src/commands/artifact/validate.ts	Adds producer-facing contract validation that runs registered schema and document-local semantic gates.
packages/prompts/src/render.ts	Renders exact contract-validation commands and the canonical coverage Markdown projection into producer prompts.
packages/runtime/src/artifact-gates.ts	Authenticates coverage evidence against production source inventory, material declarations, Recon selection, LCOV hits, goals, and Markdown publication.
packages/runtime/src/final-report-markdown.ts	Produces and verifies the canonical sanitized scoped-coverage Markdown section.
packages/evals/src/publish.ts	Carries finalized typed coverage evidence into eval publication.
packages/modal/src/public-bundle.ts	Includes validated coverage evidence and authenticated inputs in the public bundle.
packages/artifacts/schema/coverage-evidence.schema.json	Defines the strict structural producer schema paired with document-local reconciliation gates.

_{Reviews (30): Last reviewed commit: "fix(report): preserve coverage markdown ..." | Re-trigger Greptile}

[aviggiano]

Review result: hold; do not merge the current metric model.\n\nI pushed the independently valid fix that keeps nested dependency, test, and harness roots such as src/dependencies and src/test out of the production-source inventory. The remaining blocker is metric fidelity: covered is authenticated as any positive DA hit in a declaration range, so a large function with one hit and many zero-hit lines can be counted as 1/1. That is neither broad declaration completeness nor a faithful representation of the pinned covg-eval result.\n\nA direct replacement with a full-line boolean is also not safe. Pinned covg-eval has a separate identity and filtering model: it recognizes only named Solidity functions/modifiers at physical line starts, collapses overloads by name, can collide same-basename contracts, ignores constructors/fallback/receive/getters, starts after the opening-brace line, treats no body DA rows as 100%, and suppresses internal/private missing reports. The current per-declaration schema cannot reproduce that denominator exactly.\n\nThe useful path forward is to publish covg-eval output as its own evaluator-specific typed inventory/result, and publish broad production coverage as a separately named metric. Until those identities and denominators are modeled independently, the PR can still emit misleading scoped scores. I am leaving it open and unmerged.

[aviggiano]

@aviggiano ready for human review. I kept this unmerged. The coverage prompt is now concise (151 lines), with schema/tool details replaced by guidance to confirm against pinned docs or direct CLI --help. Full local CI and the replacement GitHub CI run passed, and fresh Greptile reviewed exact head bec6ca2 at 5/5 with no blocking issues.

[aviggiano]

Greptile’s current Contract validation skips reconciliation finding is stale on exact head a21aade6. ultrafuzz artifact validate first validates the immutable bytes against the registered schema, then runs executeOfflineSchemaSemanticGates and converts failures into ARTIFACT_SEMANTIC_GATE_FAILED diagnostics (packages/cli/src/commands/artifact/validate.ts:31-61). coverage-evidence.schema.json registers coverage-evidence-reconciliation (packages/artifacts/src/artifact-schema-metadata.ts:139-144), whose document-local handler calls validateCoverageEvidence. That reconciliation checks unique scopes, per-file and view totals, range/file/kind joins, selected-scope relationships, and overlaps. packages/cli/test/artifact-validate.test.ts:64-158 proves the command rejects a schema-valid but document-semantically-invalid coverage artifact; the compiled regression passes at a21aade6.

The separate zero-coverage point cannot soundly be added to offline reconciliation: covered: false means declaration-incomplete and includes partially hit ranges, whereas zero_coverage_components means no positive authenticated LCOV hit. Only runtime has the sibling LCOV authority needed for that reverse check. Greptile already agreed and withdrew that concern in #550 (comment).

No code change is warranted for either point.

[aviggiano]

@aviggiano ready for human review on exact head 812abb7b. I kept this unmerged.

The coverage prompt is now 151 lines (331 on main) and delegates schema/tool specifics to the rendered contract, pinned schemas, official docs, or direct CLI --help. I also fixed the final-report gate so canonical coverage Markdown preserves structural indentation, with regressions for indented measured output and flattened unavailable-evidence nesting.

Fresh Greptile is 5/5 with zero unresolved review threads, and all 9 exact-head checks are green, including the runtime, CLI/typecheck, package, release, Semgrep, and Socket lanes.