Public docs drift from shipped CLI, ABI v17, endpoints, registry counts, and benchmark scope

Problem

The user-facing README and usage guide disagree with current source and current keyed public records:

• README.md invokes build/examples/server, but the CMake target outputs build/examples/vllm-server.
• README and docs/USAGE.md describe ABI v10 with 19 exports, while include/vllm.h declares ABI v17 and the current public surface.
• docs/USAGE.md omits the conditional /v1/embeddings and /v1/audio/transcriptions routes implemented in api_server.cpp.
• README says 25+, 28, and 30 architectures in different places, while the registry-bound docs/FEATURES.md checker proves 35 registered and 27 gated.
• README says every architecture is token-gated, although registered scaffold/oracle-blocked rows are explicitly documented.
• README's global throughput language conflates the passing unsloth@890bdef7 27B grid with the current NVIDIA ModelOpt 27B and 35B grids that remain speed-pending.

Scope

Create one source-grounded documentation repair:

1. Correct the quickstart/server executable path.
2. Update ABI wording and the ABI evolution table through v17 from include/vllm.h.
3. Add the shipped conditional embedding and transcription endpoints to docs/USAGE.md.
4. Make README registry counts and gate wording agree with docs/FEATURES.md.
5. Scope throughput claims to the exact checkpoint and keep current ModelOpt/35B gaps honest.
6. Apply a human-language pass without adding new claims or benchmark numbers.

Do not compact docs/STATUS.md or docs/BENCHMARKS.md in this issue. Do not change code, checker semantics, lifecycle state, or accepted benchmark values.

CPU-only acceptance

• python3 scripts/check-readme-structure.py
• python3 scripts/check-public-doc-tables.py
• python3 scripts/check-supported-models.py
• python3 scripts/check-surface-coverage.py
• Configure/build the CPU examples and prove build/examples/vllm-server --help.
• Stale-string scan finds no old ABI/count/path claims.

Start with README.md and docs/USAGE.md, then verify ABI declarations in include/vllm.h, routes in api_server.cpp, and registry counts and gate wording in docs/FEATURES.md. Run the four listed Python checkers before configuring CPU examples and testing build/examples/vllm-server --help. Done means all checks pass and stale ABI, count, and path claims are absent without changing code, checker semantics, lifecycle state, or benchmark values.