fix(benchmark): gym cards name their artifact + grade on done — the 0-resolved root cause

[joelteply]

The incident (live, 2026-08-15)

Board at 0 resolved, rail banner: "16 runs gone quiet — live but no artifact activity." Citizens holding gym cards (frontier-rs / hard-rs) wander their workspaces inventing paths (swe/benchmarks/frontier-rs/src/lib.rs), and one said in-room the file she was pointed at "does not exist in this workspace."

Two defects, one seam

1. Dispatch never named the artifact. benchmark/dispatch loaded gym tasks RAW, skipping the canonical require_hands_for_code normalization cognition/eval applies at load. The card body named NO solution file (gym rows carry none), dispatch kept its own divergent {id}.rs fallback for the grading record, and eval derives sol_{id}.rs. Three surfaces, three answers — the citizen was graded against a path she was never told.

2. No grader ever fired on a gym card. benchmark_grade's done-transition arm graded SWE cards only; the "different grader" its own comment promised was never wired (gym_grader had zero done-event callers). A perfect artifact could NEVER resolve — while the kickoff promised "your artifact gets graded."

The fix (adapter doctrine holds — no parallel runner, no cognition changes)

• require_hands_for_code is now pub(crate): one derivation, both readers. Dispatch normalizes each gym task before composing the card, so the body names sol_<id>.rs and the prompt demands write→compile→verify. Title gist keeps the AUTHORED prompt so boards stay scannable and dispatch_card_key still parses.
• New gym arm in benchmark_grade::grade_card, on the SAME work/state done event the SWE arm uses: resolve the task from the same embedded gym (normalized_gym_task, same rule as the card — agreement by construction), read <home>/citizens/peers/<owner>/workspace/<solution_file>, grade through the existing test_grade harness, post the verdict into the room as a participant. Missing/empty artifact posts an ABSENCE line — never a silent skip or a fake fail.

Validation

• cargo check -p continuum-core --features metal,accelerate clean
• New regression test gym_grade_reads_the_same_artifact_path_the_card_names (pins card↔grader path agreement on a real embedded gym task) — passes
• All 9 existing commands::benchmark::tests pass (title/key contract intact)

Task #439. Live proof owed after deploy: a citizen resolving a gym card end-to-end (card names file → she writes it → done → ✅ verdict in room).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo