docs: refresh the READMEs against the work merged since May

[yeahdongcn]

The two READMEs have not kept up with the work merged since May. This PR
scans every merged PR from #61 (2026-05-07) to #128 (2026-10-09), documents
what is missing, and fixes claims that no longer match the repository. Both
languages are updated together.

Docs only: no source, test, or config change.

What was missing

Merged PR	Now documented as
#61 CUDA memory-pool aliases	"Memory management" row — torch.cuda.memory.* and the CUDA memory-pool APIs
#65 unsupported CUDA APIs → no-ops	C++ extension symbol table — cudaGridDependencySynchronize() / cudaTriggerProgrammaticLaunchCompletion()
#72 graph executable rotation	CUDA Graphs section plus the graph-rotation environment variables
#95 nested <torch/cuda.h> porting	Building C++ Extensions
#97, #104 torch.accelerator.get_memory_info()	Unified Accelerator API
#98 torch.cuda.streams	"Synchronization" row and the compatibility table
#100 stable-ABI STABLE_TORCH_LIBRARY_IMPL rekeying and stream helpers	Building C++ Extensions and the symbol table
#103 FA3 _flash_attn_forward	FlashAttention Providers section
#106 graph-safe ops, Inductor heuristics, MUSA_VISIBLE_DEVICES	torch.compile section and the compatibility table
#108 torch.cuda._get_device_index for TorchDynamo	torch.compile section
#115 generated MoE tables	Triton Fused MoE Tuning section
#121 include_paths / library_paths signatures	Building C++ Extensions
#124 FX device builtin	torch.compile section
#128 stale JIT build lock recovery	Building C++ Extensions

Structural changes

• torch_musa Compatibility (new section). Every version-gated shim was
  previously a paragraph-sized cell in "What Works". They now live in one table
  with the release each is installed on: the four < 2.11.0.post2 patches, the
  < 2.13.0 mm/bmm out_dtype= backport, asynchronous isfinite, the
  stable-ABI header backport, and torch.cuda.streams. _patch.py already
  refers to "this repository's compatibility table" — this is that table, and it
  is now readable as one.
• Environment Variables (new section). Only two of the ten user-facing
  TORCHADA_* variables were documented; the graph-rotation knobs, the
  operator-override switches, and TORCHADA_PLATFORM were only discoverable by
  reading the source.
• "What Works" goes back to one line per feature, so the table is scannable
  again.
• The verbose out_dtype risk paragraph is kept verbatim, and the remaining
  "Not covered" scope note moved with it.

Stale claim corrected

The Performance section listed the 0.1.94 / torch_musa 2.7.1 medians
(130-155 ns) and stated that all frequently-called operations complete within
200 ns. The checked-in benchmarks/benchmark_history.json already has a newer
0.1.95 entry (2026-10-09, torch_musa 2.11.0.post2+musa5.2.0) measuring
170-260 ns for the same fast paths. The section now quotes the newest
measurement, keeps the 0.1.94 figures as the comparison, and no longer claims a
sub-200 ns bound the data does not support.

Notes

• Relative links, anchors, table widths and codespell were checked; both files
  end with a newline and have no trailing whitespace.
• torchada>=0.1.95 and the mapping-rule count (400+) still match
  pyproject.toml and src/torchada/_mappings/.
• Everything added here was verified against the current tree rather than the PR
  descriptions; for example fix: generalize stable ABI extension porting for MUSA #100's description mentions CUDA_VERSION/USE_MUSA
  mappings that the merged commit does not contain, so only the rekeying and
  stream helpers it did merge are documented.

[yeahdongcn]

Merge verdict: no blocking findings — ready to merge.
Docs-only diff at 384ef94a0cf8; every new claim I could check against the tree holds, and no earlier round raised findings.

Findings: 0 blocking · 3 nit

nit	location	problem	fix
nit	README.md:456, README_CN.md:417	the 400-600ns bound is exceeded by the file it cites (606 and 626 ns medians)	quote ~400-630ns, or drop the bound
nit	README.md:441, README_CN.md:404	numbers are attributed to MTT S5000, which benchmark_history.json never records	drop the model, or add the device to the JSON
nit	README.md:196, README_CN.md:181	mirroring is not conditional on both variables being set: an unset MUSA_VISIBLE_DEVICES clears CUDA_VISIBLE_DEVICES	describe both branches

What changed · what I could not verify

Diff map. README.md +147/-23, README_CN.md +124/-21. Docs only: no source, test,
or config file is touched, so no build, import, or runtime path changes. Head commit is
384ef94a0cf8bbdccf023c95bc4c5ccf06428097 (xd/readme-refresh-2026-10), unchanged from
dispatch; mergeStateStatus: CLEAN.

Earlier rounds. gh pr view 129 --json reviews returns [] and
gh api repos/MooreThreads/torchada/pulls/129/comments returns []. Nothing earlier to
re-check; this is round 1.

Checks. gh pr checks 129 --repo MooreThreads/torchada prints no checks reported on the 'xd/readme-refresh-2026-10' branch. Per the round rules, that does not block.

Verified against the head commit (read-only; nothing was executed and no build or test
was run):

• Performance table: benchmarks/benchmark_history.json 0.1.95 (2026-10-09) medians are
  172 / 172 / 182 / 202 / 262 ns for Stream, Event, _translate_device('cuda'),
  device_count(), is_built() — matching ~170/170/180/200/260ns and the stated order.
  0.1.94 medians 121/122/136/139/159 ns support ~120-160ns; its torch_musa_version is
  2.7.1+5ee0a64.
• Environment-variable table: all ten variables exist with the documented defaults —
  TORCHADA_PLATFORM (cuda/musa/cpu), TORCHADA_EXCLUDE_DIRS,
  TORCHADA_CUDA_GRAPH_DEBUG_DUMP_PATH, TORCHADA_GRAPH_ROTATION (!= "0"),
  TORCHADA_GRAPH_EXEC_CAP (_DEFAULT_CAP = 1900), TORCHADA_GRAPH_AUTOPROBE,
  TORCHADA_GRAPH_EXEC_MARGIN (_DEFAULT_MARGIN = 128), TORCHADA_CPP_OPS_VERBOSE,
  TORCHADA_DEBUG_CPP_OPS (read in csrc/ops.h), TORCHADA_DISABLE_OP_OVERRIDE_<op>.
• Graph rotation prose: ~2048 live-executable cap, ~40 layers, template retention, LRU cap
  1900, ~0.3 ms re-instantiation, zero-cost below the cap, capture_sizes * num_layers
  instantiation — all stated in src/torchada/_graph_rotation.py.
• Compatibility table gates: _patch.py returns early from the Inductor heuristic fill-in,
  Tensor.log_, and the torch.accelerator memory overrides at
  >= "2.11.0.post2"; _cpp_ops.py:145-147 disables multinomial/log/log_ the same
  way; the func(*args, **kwargs) signature of _flash_attn_forward receives window_size
  as a 2-tuple built from window_size_left/right, matching the shim's window_size=(...)
  call. The out_dtype < 2.13.0 gate matches the >= "2.13.0" return.
• New feature rows: CUDA memory-pool API claim matches the _cuda_beginAllocateCurrentThreadToPool
  / _cuda_endAllocateToPool / _cuda_releasePool injection into torch.musa.memory;
  _get_device_index → core._utils._get_musa_device_index; get_memory_info() →
  mem_get_info; torch.cuda.streams ← torch_musa.core.stream in sys.modules.
• FA3 shim: adapter over flash_attn_func(..., return_softmax_lse=True), returns
  None, None for dropout auxiliaries, raises when output+LSE are not both returned, and is
  installed only when the provider lacks _flash_attn_forward — as documented.
• Symbol map: #include <torch/cuda.h> → torch/musa.h (cpp_extension.py:371-372);
  STABLE_TORCH_LIBRARY_IMPL(<ns>, CUDA, ...) rekey regex allows any namespace and
  whitespace layout; cudaGridDependencySynchronize() /
  cudaTriggerProgrammaticLaunchCompletion() → ((void)0) in
  _mappings/cuda_runtime.py:38-39, and they are the only two ((void)0) mappings, so
  "the one CUDA feature" holds.
• MoE recipe: --materialize and --merge-configs exist in tune_moe.py;
  tests/test_tune_moe_recipe.py has test_materialize_reproduces_the_shipped_row and
  test_checked_in_recipe_targets_the_shipped_table; ci/shapes.json and ci/models.json
  sit beside tune_moe.py; docs/tune_triton_moe.md exists.
• Links and structure: anchors #torch_musa-compatibility, #environment-variables,
  #torch_musa-兼容性, #环境变量 all resolve; every relative link resolves, including
  examples/migrate_existing_project.md#important-note-on-gpu-detection and both docs/
  targets; pyproject.toml is 0.1.95; src/torchada/_mappings/ holds ~422 rule lines, so
  400+ holds; the decode_jpegs_cuda.h example is the one in the porter's own comment;
  both files end with a newline and have no trailing whitespace; code fences pair up.

Assumed, not verified. That the 0.1.94/0.1.95 runs were on MTT S5000 — the JSON records
only "platform": "MUSA", which is the basis for nit 2. I did not execute anything, did not
re-measure, and did not check the vendor claim that 2.13.0 fixes out_dtype (the PR keeps
that framing as an accepted risk, matching the source comments).

auto-review round 1/3 · verdict: clean