[Issacboy123]

Summary

Evolution could silently skip unreadable trace evidence by advancing its processed cursor past a failed materialization. This change stops processing at the first failure, keeps that trace eligible for retry, and exposes a generic retry notice while still allowing a valid earlier prefix to be processed.

This PR is stacked on the Together model-ID correction.

Reasoning

Representing the first materialization failure at the evidence-catalog boundary gives the control plane a precise contiguous range to commit. This prevents manual and automatic jobs from consuming later evidence out of order. Successful materializations remain reusable while pending, then are discarded after their cursor is committed, avoiding indefinite retention without caching failures.

Changes

• Replaced the successes-only evidence query with an ordered scan that returns the successful prefix and generic metadata for the first failed trace.
• Prevented manual jobs, automatic eligibility, and excluded-only cursor commits from advancing beyond the first failed trace.
• Retried failed materializations on later scans and discarded successful cached entries after their cursors were committed.
• Added the retry notice to the evolution overview and interactive menu without exposing exception text.
• Added regressions for cache reuse and cleanup, retry behavior, safe cursor ranges, threshold truncation, exclusions, overview counts, and menu rendering.

Testing

• node ../../node_modules/vitest/dist/cli.js --run test/evolution-runtime.test.ts test/evolution-menu.test.ts
• node ../../node_modules/vitest/dist/cli.js --run test/invocation-tracing.test.ts
• node ../../node_modules/vitest/dist/cli.js --run test/control-plane.test.ts test/control-service.test.ts
• npm run check

[Pepps233]

Add a maximum retry amount / retry counter per unreadable / broken trace. Currently, if a trace is broken due to permission issues such as file permissions, it will retry forever. Maybe add a maximum retry amount of 2?