[edsiper]

A successful heap benchmark batch left freed pointers in roots. If a later batch failed before replacing every entry, cleanup could destroy those pointers again. Clear each entry after destruction and add a regression test that injects failure at the first, middle, and last record of a later batch.

Clarify that the large-object threshold applies only to arena-aware owned SDS payloads larger than 1 KiB. Payloads up to and including 1 KiB use pooled size classes regardless of the threshold; the existing size-class and cfl_sds_alloc() capacity descriptions are preserved.

Validation:

• BUILD_DIR=/tmp/cfl-feedback-verify scripts/agent-verify.sh: all 35 tests passed; whitespace checks passed.
• valgrind --quiet --leak-check=full --show-leak-kinds=all --errors-for-leak-kinds=all --error-exitcode=1 /tmp/cfl-feedback-verify/tests/cfl-test-benchmark_records: passed without errors or leaks.
• The same regression test against a temporary copy of the original benchmark reproduced a use-after-free under Valgrind.
• Heap and arena benchmark smoke checks passed with 8 3 owned 1.

Only benchmark behavior, its test registration, and documentation change. No library implementation, API, ABI, or allocator behavior changes; consumers do not need an upgrade for this fix. Cross-platform and downstream checks were not run locally and remain CI coverage.