Make HARP scoring deterministic and behaviorally valid #6

Problem

The scorer does not consistently measure its stated rubric. Phase 1 treats the first skill_view anywhere as success rather than requiring it as the first tool action; skill matching is overly broad; and Phase 3 infers context preservation and completion quality from message/tool counts and response length. These shortcuts permit false passes and make readiness verdicts indefensible.

Goal

Replace heuristic scoring with deterministic, probe-specific assertions over ordered tool traces and produced artifacts.

Scope

• Require exact expected tool/action ordering where a probe specifies it.
• Replace permissive skill matching with explicit aliases declared by each probe.
• Define structured assertions for files, tool arguments/results, and final deliverables.
• Verify Phase 3 context transfer through facts/citations extracted from earlier outputs, not transcript length.
• Emit per-assertion pass/fail evidence and diagnostics in the score JSON.
• Add scorer unit tests with committed positive, negative, and adversarial session fixtures.

Acceptance Criteria

• Every rubric rule maps to an executable assertion or is removed/reworded.
• A session that calls a raw tool before the correct skill_view fails the Phase 1 first-action criterion.
• Wrong skills sharing a generic path segment cannot pass matching.
• Out-of-order chains, missing artifact creation, and unrelated verbose final answers fail the relevant Phase 3 assertions.
• Test fixtures demonstrate each former false-positive class and pass in CI.
• Score output identifies the exact failed assertion and supporting trace/artifact.

Non-Goals

• LLM-as-a-judge evaluation of subjective prose quality.
• Expanding the number of public probes.

Implementation Notes

Keep assertions data-driven in the probe definitions where practical. Parse tool-call arguments and results defensively rather than relying on preview strings.

Dependencies

• Depends on Restore a self-contained, runnable benchmark pipeline #5 for a stable probe manifest and fixture/report layout.