Divergent benchmark can omit requested worktrees and truncate origin seed names #1522

Problem

Source-revalidated on main 3f2a835c78e9daba7baefc4c018c6fc9ca3c6e4a.

Two original benchmark contract defects remain. The runner validates a worker minimum and worktree multiple-of-four, but not workers>=worktrees; each worker gets one fixed worktree and verify groups successes only by kind. initial_sync_unbounded separately derives the base with split_once("--wt-"), truncating a legitimate origin name containing that delimiter.

Location: crates/prod-code-client/src/divergent_bench/runner.rs:27-42,108-110; crates/prod-code-client/src/divergent_bench/session.rs:122-134.

Reproduction boundary

For --worktrees 16 and default12 workers, copies12-15 are never queried yet earlier copies cover every kind and verification can pass. For a base repository named api--wt-tools, initial sync derives base api instead of api--wt-tools-divergent-bench, defeating the intended warm seed. These are source deductions; no benchmark or performance numbers were produced.

Expected correction

Require coverage of every requested worktree (or reject insufficient workers), retaining per-copy evidence rather than kind-only success. Derive the seed from the known origin identity or only the generated terminal suffix. Verify these bounded setup/scheduling cases without launching a broad benchmark campaign.

Verification limits

Read-only current source inspection; affected files unchanged from the audit checkout. No new tests, build, benchmark, LSP session, server restart or platform run.

Original review findings

• divergent-bench --worktrees N: measure 16 concurrent agent worktrees (#406) #407 (comment)
• perf: bound concurrent cold Rust loads and seed benchmark copies #432 (comment)