fix(dashboard): dict belongs in Stage, not as fake levels + LDM level labels

[polaz]

Problem

The LDM+dict bench matrix bakes the _dict discriminator into the level name (level_1_fast_ldm_dict, level_22_btultra2_ldm_dict in zstd/benches/support/mod.rs). That name leaks into the level field of dashboard records, causing three regressions on the Level Profile:

1. Fake levels in the Level dropdown — level_1_fast_ldm_dict / level_22_btultra2_ldm_dict appear as separate levels. They should not exist as levels; the dict/plain split must be conveyed by the Stage dropdown (compress-dict / decompress-dict, which already exist).
2. Wrong level labels — parseLevelId only matches level_<N>_<strategy>, so the _ldm / _ldm_dict variants fail to parse and render as raw IDs sorted to the bottom.
3. "Staircase" chart — PR ci(bench): fix dashboard dict ratio/labelling + wasm-pack binaryen flake #376 split plain/dict into distinct (level, kind) X-axis points, doubling every level column. Dict should instead be separate toggleable series (Rust/FFI compress-dict + decompress-dict) on the same real-level X-axis.

Fix

• CI parser (run-benchmarks.sh): normalize level (strip the _dict bench-variant suffix) so the dashboard level axis carries only real level identity; route dict-ness entirely through stage. Memory rows derive their dict stage (compress-dict/decompress-dict) from the stripped suffix. No Rust bench changes (keeps the CI level-filter inventory intact).
• Dashboard: revert the (level, kind) X-axis doubling; X-axis = real levels. Add 4 dict series (Rust/FFI x compress-dict/decompress-dict) with checkboxes, kept separate from plain via stage-kind bucketing. Teach parseLevelId the _ldm suffix. Keep history mode kind-aware; keep the aggregate chart plain-only.

Acceptance

• Level dropdown shows only real levels (level_1_fast_ldm, level_22_btultra2_ldm, numeric, na), nicely labelled.
• Stage dropdown distinguishes compress/decompress/compress-dict/decompress-dict.
• Level Profile: real-level X-axis, 10 toggleable series (6 plain + 4 dict), no staircase.

Estimate: 3h

[coderabbitai]

🔗 Related PRs

#142 - ci(bench): cover full -7..=-1, 1..=22 level range with strategy-tag labels [merged]
#193 - feat(dashboard): level-profile chart (6 lines) + HTML legend + sign-gated parity band [merged]
#236 - test(bench): add decompress-dict group to compare_ffi [merged]
#277 - test(bench): unblock dict-driven bench matrix + add pure_rust_with_dict compress arm [merged]
#376 - ci(bench): fix dashboard dict ratio/labelling + wasm-pack binaryen flake [closed]

---

📝 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!