fix(results): hard-remove legacy grading.json compat reads in manifest.ts

[christso]

Summary

Found while auditing av-kfik.46.4/av-kfik.46.5 (result-artifact contract cleanup). apps/cli/src/commands/results/manifest.ts still tolerantly read a pre-major-version grading.json shape:

• assertion_results as a fallback for assertion detail
• verdict === 'pass' / score-threshold guessing as a fallback for deriving pass
• graders/evaluators as legacy aliases for component_results — the evaluators branch was explicitly commented "evaluators was renamed to graders in v4.13 — read both for backwards compat... TODO: remove... once old run directories are no longer in use."

Meanwhile apps/cli/src/commands/results/validate.ts already treats all of these same field names (assertion_results, assertions, passed, evidence, verdict, graders, checks) as hard validation errors. One code path rejected the legacy shape as invalid; another silently read it. Since this major version has no production users yet, there's no on-disk bundle to preserve compatibility for — hard-removing is correct rather than adding test coverage for a fallback that will never be exercised.

Changes

• Removed readGradingAssertionResults()/mapGradingAssertions() — the assertion_results fallback. Call site now derives directly from the grading component itself (mapComponentAssertion), matching what already happens when component_results is present.
• Removed the graders/evaluators branches in readNestedGradingScores() and the ad hoc type augmentation in hydrateManifestRecord() — both now read only component_results (and the separate, still-current scores field).
• Simplified mapGradingEvaluator()'s pass derivation to evaluator.pass === true — pass is always present per the current grading.json contract, so the verdict/score-threshold fallback was dead weight.

Dashboard's EvalDetail.test.ts already has a test asserting these exact legacy fields (assertion_results, evidence, graders, checks) are ignored when parsing a grading artifact — that code path (parseGradingArtifact/parseComponentResult) was already written correctly with no legacy tolerance, so no changes needed there. No test anywhere exercised manifest.ts's now-removed fallback branches (confirmed via search) — nothing to update/remove on the test side beyond the source itself.

Test plan

• bun run typecheck (workspace) — clean.
• bunx biome check on the touched file — clean.
• bun test apps/cli/test/commands/results — 252/252 pass.
• bun test apps/dashboard/src/components/EvalDetail.test.ts — 5/5 pass (unaffected, confirms Dashboard-side parsing was already correct).
• Live check: ran a real eval (examples/features/rubric/evals/operators.eval.yaml, live Azure target + grader), then results summary/results show against the resulting bundle — both render correctly through the simplified code path (assertions with text/passed/evidence, correct scores).

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	b33c2d3
Status:	✅  Deploy successful!
Preview URL:	https://900ed169.agentv.pages.dev
Branch Preview URL:	https://fix-hard-remove-legacy-gradi.agentv.pages.dev

View logs