docs(design): intra-file parallel rolling-hash for large files (#2206)

[oferchen]

Summary

• Adds docs/design/delta-intra-file-parallel.md capturing the design space for splitting a single large basis file (>= ~100 MB) into overlapping windows and running N parallel rolling-hash matchers, with the merged token stream emitted in input-offset order.
• Cites the current sequential matcher (crates/matching/src/generator.rs:141 DeltaGenerator::generate, hot loop at :196-263) and its sole production caller (crates/transfer/src/generator/delta.rs:171).
• Proposes MIN_PARALLEL_FILE_SIZE_BYTES = 128 * 1024 * 1024 next to existing thresholds in crates/transfer/src/parallel_io.rs, with rationale tied to calculate_block_length at crates/signature/src/block_size.rs:129.
• Cross-references the upstream match.c reference doc (docs(design): upstream match.c and token.c reference for delta work (#1580) #4202), MD4/MD5 multibuf SIMD benches (perf(checksums): bench AVX2 multi-buffer MD4 scalar vs batch (#2100) #4189, bench(checksums): MD5 scalar vs batch throughput (#2101) #4191), zsync adversarial fixtures (Respect last occurrence for tri-state toggles #2079, Ensure delete-after uses content verification before skipping copy #2080), zsync benches (Ensure identical timestamp files are compared by content #2063, Ensure --ignore-times forces a rewrite #2067, Ensure copy skip heuristics verify file content #2071), and the sequence-indexed ReorderBuffer work (Ensure release task uploads cross-compiled binaries #1885).

Recommendation

Defer. Do not implement intra-file parallel rolling-hash until the MD4/MD5 multibuf SIMD benches (#4189, #4191) show single-thread strong-checksum verify is no longer the hot stage on representative workloads. Either lifts single-thread throughput enough that intra-file parallelism becomes unnecessary, or it does not - the decision must rest on bench numbers, not design speculation. A prototype-behind-feature path is rejected because behavioural divergence between gated and default builds is exactly what the golden-byte parity tests are designed to catch.

Wire-format

No wire-format change. Token stream MUST be byte-identical to the serial matcher's output for every input. crates/protocol/tests/golden/ and the interop matrix must remain green for any future implementation PR.

Test plan

• Pure docs change - no code touched.
• cargo fmt --all clean (verified before push).
• Doc has no AI tooling references, no em-dashes.
• All file:line citations resolve in the current tree.