docs(design): defer ABW-2/3/4 pending BR-3j.f bench evidence (ABW-1 audit closure)

[oferchen]

Summary

• Adds docs/design/abw-2-pipelined-verify-write-deferred-2026-05-21.md, a closure note that defers ABW-2 (test: add tests for SHORT_SUM_LENGTH vs SUM_LENGTH phase toggle #2571), ABW-3 (fix: symlink itemize never reports size changes #2572), and ABW-4 (fix: wire --ignore-existing, --existing, --max-delete into server path #2573) pending BR-3j.f (feat(daemon): add max verbosity parsing and enforcement #2508) re-bench evidence.
• Discharges the ABW-1 audit recommendation (docs/audits/abw-1-apply-batch-verify-write-overlap-2026-05-21.md, PR docs(audit): apply_batch_parallel verify-vs-write overlap potential (ABW-1) #4670, section 4) to skip the pipelined verify/write design until bench data shows 0.5 <= verify_wall/write_wall <= 2.0 on a production-relevant workload cell.
• No source edits; pure markdown, 166 lines, with file:line cites into crates/engine/src/concurrent_delta/parallel_apply.rs for apply_batch_parallel (515-542), FileSlot::ingest (248-258), and verify_chunk (632-652).

Why defer

• ABW-1 quantified the pipelining ceiling: 1.13x balanced, 1.50x CPU-bound verify, 1.03x I/O-bound write, ~0x single-file.
• apply_batch_parallel has zero production callers today (per RJN-1, PR docs(audits): apply_chunk_parallel call sites and per-chunk dispatch benefit (RJN-1) #4656). The PIP-3+5 dispatch heuristic (PR perf(transfer): enable parallel receive-delta by default via Path B heuristic (PIP-3 + PIP-5) #4666) gates single-file batches out of parallel-receive-delta entirely, so the degenerate case is already off the hot path.
• A ~200-300 LoC bounded-channel writer + backpressure + error-propagation rework is not worth shipping for a 1.13x balanced-case win on a path no production caller reaches.

What unblocks the work

BR-3j.f (#2508) extends crates/engine/benches/parallel_receive_delta_perf.rs to emit verify_wall and write_wall per workload cell. The doc lifts the audit's decision gate verbatim: ratio in [0.5, 2.0] proceeds to ABW-2; outside that range closes out the line of work in project_apply_batch_write_serial.md.

Closure shape

• ABW-2/3/4 stay open in the tracker, labelled deferred pending BR-3j.f bench evidence.
• Per-file Mutex<FileSlot> is named as the real bottleneck; a future multi-threaded-per-file writer or a CPU-bound verify regime would re-open ABW-2. Both flagged as out-of-scope future options so contributors with bench data know which assumptions this closure rests on.

Test plan

• cargo fmt --all (no source files changed)
• CI required checks pass (fmt+clippy, nextest stable, Windows, macOS, Linux musl)