v2.9.8

v2.9.8 (2026-09-05)

Add long-task boundary preservation coverage to the agentic benchmark

• Add the matrix-v7 long-task-boundary-preservation scenario with development, held-out-normal, and held-out-boundary cases (33 cases across 11 classes), including an indirect restock consumer in the normal case so the attribute rename is exercised beyond the partner export. The normal case is a sentinel gate rather than a discriminator, while scoring stays arm-neutral and the benchmark's advisory evidence boundary is preserved.
• Keep the three committed pre-v7 public snapshots available for exact, verification-only projection without relabeling them as current evidence.

Verification boundary

• Version-bump audit passed: 8 declared manifests in sync at 2.9.8; no undeclared references.
• PR #44 CI (method-pack-checks, CodeRabbit review) passed on the merged head; post-merge validation on main passed: case manifest (33 cases, 11 scenarios), matrix validation, renderer self-test (6 profile goldens, 3 historical snapshots, 46 negative cases); git diff --check passed.
• No live held-out measurement of the new boundary discriminator on the final commit; the three CodeRabbit findings were verified fixed by reviewer code inspection.