fix(expert-skill): align FlyDSL selectors and validation

[IzacharyI]

Summary

Follow-up to #386 (review) addressing the post-merge review findings :

• Fix expert-skill selector aliases and emitted workflow modes so the
  gfx950 FlyDSL MoE and dense GEMM skills can be reached correctly.
• Clarify that the three submitted skills are validated by isolated
  kernel A/B and parity; this PR does not claim new E2E gains.
• Add L0 CI coverage, static skill validation, and structured dense
  blockscale evidence with per-run latencies.

Changes

• Make the two FlyDSL MoE skills match both legal operator identities:
  • grouped_gemm_moe
  • fused_moe_grouped_gemm
• Make the gfx950 blockscale GEMM skill match:
  • dense_gemm
  • scaled_quant_gemm
• Define scalar-or-list operator matching in the kernel/e2e expert-skill guidance.
• Mark the MoE recipes as FlyDSL in-place optimizations (flydsl -> flydsl).
• Fix --emit-plan:
  • same-backend recipes use mode=optimize
  • cross-backend ports use mode=author target_language=...
  • mixed-source skills expose both applicable arms
• Remove unenforced e2e_delta_min_pct fields from the three kernel-scope skills. Their validated status now clearly represents isolated kernel A/B + parity only.
• Add structured gfx950 dense evidence with three archived on-box CK/FlyDSL latency pairs:
  • median CK: 0.187461 ms
  • median FlyDSL: 0.128937 ms
  • ratio-of-medians: 1.4539x
  • Director-audited pair: 0.187491 -> 0.128937 ms = 1.4541x
• Correct the down-proj tuning guidance:
  • tile_n=256/tile_k=128 is the decisive configuration
  • XCD swizzle and fused-promote were non-load-bearing for this shape
  • the 8-wave follow-up was a measured dead-end
• Clarify that use_expert_skills=false suppresses the registry fragment, while general target-backend knowledge remains in the base role prompts.
• Add the portability regression test and three static skill validations to L0 CI.
• Preserve ensure_flydsl as a non-profile-matched dependency skill after reindexing.
• Remove the non-canonical ck -> ck_tile example.

Validation Scope

The three submitted skills are scope: kernel.

Validated isolated results:

• Decode MoE stage-1 segment: 1.1043x (-9.45%)
• Prefill MoE stage-2 segment: 1.2099x (-17.35%)
• Dense blockscale down-proj: 1.4539x ratio-of-medians; Director pair 1.4541x

These numbers include parity validation and the complete declared kernel segment.

This PR does not claim a newly validated serving-level throughput, TTFT, or TPOT improvement. E2E acceptance remains the responsibility of the normal serving A/B gate.

Evidence

• flydsl_decode_moe_stage1_blkmap/validation_flydsl_0_2_2.yaml
• flydsl_prefill_moe_stage2_fp8partial/validation_flydsl_0_2_2.yaml
• flydsl_fp8_blockscale_gemm/validation_gfx950.yaml

The dense YAML was created from existing on-box GEAK measurements. The original raw reports remain external, and the exact FlyDSL/AITER commits were not captured in the original run.

Validation

• 19 selector/emit-plan/evidence regression tests pass.
• Static validation passes for all three skills.
• index.yaml regenerates without changes.
• Dense evidence records three archived on-box CK/FlyDSL latency pairs.
• Serving-level throughput, TTFT, and TPOT were not revalidated in this PR.