test(bench): add dictionary training + dict-driven compress/decompress benches vs FFI

[polaz]

Context

Current compare_ffi bench suite covers compress + decompress vs FFI across multiple corpora and levels, but does NOT exercise the dictionary-builder path (fastcover / cover) nor compress/decompress with a pre-trained dictionary. Dictionary-driven workflows are a major real-world use case (database column compression, log shipping, RPC payloads) where our Rust implementation must hold up against libzstd on both:

1. Dictionary training throughput — ZSTD_trainFromBuffer_fastCover vs our dict_builder feature
2. Compress/decompress with dictionary — every level matrix, vs FFI's ZSTD_compress_usingCDict / ZSTD_decompress_usingDDict

Problem

Without coverage, regressions in either path land silently. The bench-dashboard tracks compress/decompress deltas but the dictionary track is invisible — a 30% regression in dict training would never be flagged.

Proposed scope

Bench additions

• bench_dict_training in zstd/benches/compare_ffi.rs:
  • Train dict on each corpus (sample budget = realistic ~1 MiB pool, dict size = 32 KiB)
  • Two sides: pure Rust dict_builder::fastcover vs FFI ZSTD_trainFromBuffer_fastCover
  • Throughput metric: input bytes per second (training is bound by sample size, not output size)
• bench_compress_with_dict and bench_decompress_with_dict:
  • For every level in the existing compress matrix
  • Train dict once per corpus (re-use across levels)
  • Compress and decompress with that dict, both sides
  • Existing per-stage report-line shape

Dashboard wiring

• New stage labels: dict_train, compress_with_dict, decompress_with_dict
• benchmark-relative.json records the same (rust_value, ffi_value, delta_ratio) triple per (level, scenario, target) for these stages
• Dashboard's Level Profile + Aggregate chart pick them up via existing profileLogicalMetric after extending its switch

Acceptance criteria

• CI bench job runs the new benches without regression in compute time (cap dict-training measurement at ≤ 30 s per scenario)
• benchmark-relative.json carries the new stages on every snapshot
• Dashboard Aggregate chart shows dict-train ratio and dict-driven compress/decompress alongside the existing series
• Sanity: rust dict-train delta vs FFI lands within reasonable parity band (initial measurement, no SLO yet)

Files involved

• zstd/benches/compare_ffi.rs — add the three new bench groups
• .github/scripts/run-benchmarks.sh — emit JSON rows for the new stages
• .github/bench-dashboard/index.html — extend profileLogicalMetric to map the new stages

Out of scope

• Custom cover algorithm bench (only fastcover for now — that's what we ship)
• EE-only dict policies (this is CE-side parity work)

[coderabbitai]

🔗 Related PRs

#38 - test(bench): expand zstd benchmark suite [merged]
#43 - test(bench): expand benchmark parity matrix [merged]
#73 - perf(bench): add fastcover vs ffi dict-training delta [merged]
#143 - bench: add real peak-alloc metric for full encode/decode/memory baseline [merged]
#149 - feat(bench): #99 Rust↔FFI sequence-stream comparator [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!