[Longxmas]

Keep one current, reproducible result set per model and device. Remove measurements that newer runs replaced, and notes that only recorded how a result was produced. H200 numbers are not kept in the published benchmark index.

Summary

• Replace the September 7 aggregate in docs/{en,zh}/benchmark.md with a latest-results table (one row per model × device), refresh coverage gaps, and drop the H200 engine-mechanism section and the October 14 prefix GELU note.
• Remove superseded pi0.5 Thor / RTX 4090 quantization replays and early 1,600-frame rows; keep AGX Orin quant results and add the Orin tuned-kernel latency row.
• Delete the August 3B-navigation evaluation reports; latency for those models now points at qwenvl-benchmark.
• Condense ActiveVLN, GR00T, OFT, LingBot, DM0.5, Cosmos, StreamVLN, Qwen and related READMEs to results, conditions and commands (no rejected candidates, local reports/ paths, snapshot tars or checksum lists).
• Rename VLN device sections by protocol/date; sync VLN and RLinf tables into README.zh-CN.md; fix related anchors in docs/{en,zh}/api.md.

Type of change

• Bug fix
• New model adapter or policy-contract change
• Engine, scheduler, session, or rollout change
• Kernel, layer, or numerical change
• Documentation
• Build, dependency group, or lock-file change

Verification

Docs-only change; no CUDA or weight runs.

$ python scripts/check_commit_policy.py
# passed for d60e49b

$ mkdocs build --strict -f mkdocs.yml   # en
$ mkdocs build --strict -f mkdocs.zh.yml  # zh
# both passed (Windows: temp copy / symlink workaround)

$ git diff --check
# clean

• Hardware: not applicable (documentation only)
• Weights: none
• Parity: not applicable (no numerical re-measurement in this PR; published numbers are the retained latest runs already on the branch)

Engine invariants

• No model-name branch was added to the engine; capabilities are declared
  through ActionDecoder / RLDecoder and the shared metadata.
• Parity tests under the affected layers or backend still pass, or the
  intentional difference is documented with numbers.
• New dependencies are declared in the correct profile group, and
  uv.lock is consistent.
• No deployment-runtime code was imported across the boundary; integrations
  use the versioned API.

Checklist

• My PR title and every introduced commit follow the
  commit policy,
  and every author/coauthor has supplied their own DCO Signed-off-by.
• New models follow the flow in
  CONTRIBUTING.md
  and have a design document under docs/proposals/ (N/A: no new model).
• Documentation reflects the change, and parity or runtime numbers in
  docs/en/benchmark.md
  are updated when they change.
• No checkpoints, datasets, benchmark outputs, recordings, credentials,
  addresses, private paths, or machine-specific scripts are included.
• I agree that this contribution is licensed under Apache-2.0, per
  CONTRIBUTING.md.