perf(benchmark): bound terminal run/batch registries

[beruro]

Problem

The process-wide benchmark run and batch registries retained every terminal entry for the lifetime of the app. Repeated benchmark use therefore grew memory without a bound, and evaluator subprocesses could remain alive when the app exited.

Solution

• Keep every active run/batch resident, but retain only the 50 most recent terminal entries in each registry.
• Apply a pure, deterministic retention policy whenever an entry is inserted or transitions to a terminal state; no polling or sweep timer was added.
• Preserve persisted batch recovery: evicted batches continue to reload from batch history on demand.
• Best-effort terminate running evaluator processes from the existing app-exit lifecycle next to PTY cleanup.
• Unit-test active-entry protection, ordering, tie breaks, zero budgets, idempotency, status classification, and both registry adapters.

The branch is merged with the current develop; GitHub reports it mergeable.

Potential risks

• Terminal standalone runs older than the newest 50 are no longer queryable because that registry has no persistence layer. Active runs are never evicted, and persisted batch results remain reloadable.
• A run permanently stuck in an active state remains resident. This is conservative and matches the prior lifecycle behavior.
• A pre-existing batch persistence failure can still lose the latest state after eviction; persistence failures are logged but are not made transactional by this PR.
• Exit cleanup uses try_lock and SIGTERM, so it is best-effort and cannot cover a hard process kill.
• There is no schema, DTO, IPC, or wire-format change. Rollback is a normal revert, which restores unbounded in-memory retention and the previous exit behavior.

Audit

Architecture review covered registry ownership, status/FSM classification, persistence fallback, lock boundaries, process lifecycle, and IPC compatibility. Performance-guard verdict: pass — terminal memory is capped, active work is not evicted, pruning runs only at existing mutations, no idle/background loop was introduced, and reopen/reload uses the existing persisted batch path.

Verification

• CARGO_TARGET_DIR=/Users/junyu/github/ORGII/src-tauri/target cargo test -p org2 benchmark::retention — PASS, 11 passed / 0 failed.
• rustfmt on the touched Rust files — PASS.
• git diff --check — PASS.
• Merge of current origin/develop — PASS, no conflicts; targeted tests were run after integration.