perf(encode): close the L6-L9 Row/lazy dictionary speed gap

[polaz]

Problem

After the small-frame FSE-table-selection fix (PR #433), Fast/dfast dictionary frames are at or below libzstd parity on both arches. The remaining dictionary-path gap is at the Row / lazy backends (levels ~6-9), which stay 2-3x slower than libzstd and also lose a little ratio.

Authoritative compare numbers (ours vs libzstd compress_usingCDict, same host, dict_matrix example, raw-content 8 KiB dict, dict-id flag off):

fixture	lvl	i9 (x86) sp	M1 (aarch64) sp	ratio
logs-1k	6	2.75x	2.10x	1.44x (26 vs 18 B)
logs-1k	9	2.02x	2.00x	1.00x
logs-4k	9	2.05x	1.94x	1.00x
logs-16k	9	1.80x	1.86x	1.00x

(sp = ours_us / c_us; >1 = slower. ratio = our_bytes / c_bytes.)

Scope

• Profile the L6 (Row/greedy-lazy) and L9 (HashChain/lazy2) dictionary path on both arches (x86 BMI2 on the i9, NEON on the M1 — a same-hardware ours-vs-libzstd delta is authoritative per arch).
• L6 also shows a ratio gap (we emit 26 vs libzstd's 18 bytes on logs-1k): the lazy matcher finds fewer/shorter dictionary matches on the dict-primed chains. Ratio is the first priority.
• Likely targets: the Row tag-mask SIMD coverage, dict-primed chain traversal depth, and the lazy carry-forward decision on dict matches.

Acceptance

• L6/L9 dictionary throughput within ~1.2x of libzstd on the bench fixtures (both arches), ratio >= libzstd.
• No regression on the Fast/dfast/BT dictionary paths or non-dict levels.

Follow-up to PR #433 (which fixed the 5.5x copy-mode Fast dict ratio bug, the resident-dict retained-budget over-window bug, and the small-frame FSE-selection hotspot).

[coderabbitai]

🔗 Related PRs

#277 - test(bench): unblock dict-driven bench matrix + add pure_rust_with_dict compress arm [merged]
#328 - perf(encode): size-gated dictionary match-finding for the Fast strategy [merged]
#331 - perf(encode): close level-5 greedy speed gap (row match-span cap) [merged]
#390 - perf(encode): reshape dfast match-find toward donor structure [merged]
#426 - perf(encode): close small-window dict compress gap (hash-chain matchfinder + borrowed dict kernel) [merged]
#427 - perf(hc): lazy-skip + separate dictMatchState beat C on small-window dict compress [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!