fix: repair six stale documented commands and add a check so they stop rotting

[justinchuby]

Answering "is everything broken fixed?" properly meant checking rather than
recalling, so I validated every cargo command written down in the repo against
cargo metadata. 262 commands; six named something that does not exist.

Two were my own fallout from the feature rename (#1629), which I missed because
I updated .github/skills/ and not .agents/skills/:

• .agents/skills/profiling/SKILL.md still asked for bench-native,cuda.
• crates/onnx-genai-capi/README.md still asked for --features cuda.

One is a shell script with exactly the two bugs I fixed in #1630, which I should
have found then by looking at the class of defect instead of the instance:

• scripts/build_real_model.sh built -p onnx-genai --bin onnx-genai (that
  package ships only the diffusion bins) and called generate --model DIR (the
  model directory is positional).

Three predate all of this and are worse than stale -- they were never true:

• docs/ep-plugin/EP_PLUGIN_EXPORT_TEST_PLAN.md built a plugin-export feature
  on onnx-runtime-ep-cpu and grepped for CreateEpApiFactories. What shipped
  is a separate onnx-runtime-ep-cpu-plugin cdylib exporting CreateEpFactories
  and ReleaseEpFactory. Verified by building it and reading nm -D.
• docs/performance/CPU_MATMUL_ASSIGNMENT.md gave a repro recipe using a
  bench_prec binary with --native-threads / --ort-intra-threads. git log
  finds no commit that ever added or removed it. The recipe is removed rather
  than annotated, because a command that cannot run is not worth keeping; the
  section now says plainly that those numbers cannot be reproduced as written.
• A fixture generator credited its canonical output to
  cargo run -p onnx-std --example convert_fixture. That example has never
  existed either.

The reason all six survived is that nothing ever checks a command that nothing
ever runs; the failure then looks like the reader's environment rather than the
line. scripts/check_documented_commands.py now validates the parts that can be
checked statically -- package, --bin, --example, --test, --bench,
--features -- and runs in the rust-quality lane. Dated docs and .squad
archives are skipped: they record what was run at the time.

Verification:

• The check reports 262 commands, all resolving, on this branch.
• It has teeth, by the same --self-test convention the dispatch-manifest lint
  uses: six cases, five that must be detected (bad feature, bad bin, unknown
  package, bad bench, bad dep/feature) and one valid command that must stay
  silent. Also confirmed by hand against the real tree, by breaking
  docs/build-features.md and the profiling skill and watching each get caught.
• bash -n clean on the repaired script, and its new CLI form reaches model
  loading instead of being rejected by clap.
• ci.yml parses as YAML.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com
Copilot-Session: 0190e2eb-abe4-451f-b36d-44a035a99b7e

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.49%. Comparing base (f8eb8a3) to head (1b2c8a7).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1635      +/-   ##
==========================================
- Coverage   81.63%   81.49%   -0.15%     
==========================================
  Files         384      384              
  Lines      180249   180249              
  Branches   180249   180249              
==========================================
- Hits       147150   146885     -265     
- Misses      28155    28418     +263     
- Partials     4944     4946       +2     

Flag	Coverage Δ
cli-ort-linux	82.60% <ø> (ø)
cli-ort-windows	82.20% <ø> (+0.09%)	⬆️
mlas	85.09% <ø> (ø)
offline	81.38% <ø> (-0.16%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 4 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.