Reject vacuous standardized coverage outputs even when tooling exits zero #590

Summary

Ultrafuzz should machine-reject standardized coverage results with an empty target map, a zero denominator, or no production-attributed sources, even when the coverage tools exit successfully.

The v0.0.16 coverage schemas can represent a typed blocker and reject an empty reported metric list, but the runtime does not validate the semantic contents of the underlying recon-generate coverage and covg-eval outputs. Correct rejection still depends on the model noticing that the result is vacuous.

Evidence from #531

The Aave v4 stateful-invariant-coverage node observed:

• recon-generate coverage exited 0 and wrote a two-byte {} Magic coverage map.
• covg-eval magic/ echidna/ --return-json exited 0.
• The result reported full_coverage: true while functions_analyzed was 0.
• The agent correctly rejected the result, recorded coverage-tooling-blocked, and left the standardized percentage null.

That was a good model decision, not a runtime guarantee. Another producer could turn the same successful exit codes or vacuous "full coverage" result into a reported percentage.

Expected behavior

Coverage acceptance must depend on a nonempty, production-attributed denominator rather than process exit status or a tool's boolean "full coverage" field.

The node may complete successfully with a typed coverage blocker; this issue does not require a tooling problem to fail the entire workflow. It requires the unusable measurement to be impossible to publish as valid coverage.

Acceptance criteria

• Capture and validate the exact standardized-coverage tool outputs as typed durable evidence.
• Reject an empty coverage target/map such as {}, regardless of exit code.
• Reject functions_analyzed == 0 or an equivalent zero denominator.
• Reject a result with no source attributable to configured production roots.
• A rejected result cannot use target-met or below-target and cannot publish a coverage percentage.
• Record a typed coverage-tooling-blocked or production-source-attribution-blocked outcome with the raw evidence paths.
• Propagate the blocker into campaign/report comparability.
• Add regressions for: empty map plus exit 0; zero analyzed functions plus "full coverage"; harness-only sources; and a valid nonzero production-attributed result.
• Keep the gate adapter-friendly so coverage backends other than Recon can supply the same denominator and source-attribution facts.

Related: #531 and #532.