docs: align public surface with shipped interfaces

[localai-org-maint-bot]

Summary

• fix the README quickstart to invoke the shipped vllm-server binary
• update README and usage docs from ABI v10 to the current ABI v17 surface
• document the conditional embeddings and audio-transcription routes
• distinguish 35 registered architectures from the 27 correctness-gated text-generation architectures
• scope the passing Qwen3.6-27B throughput claim and badge to unsloth revision 890bdef7, while keeping NVIDIA ModelOpt 27B and current 35B gaps visible
• add the durable positioning rule that public claims, including badges, distinguish registered from gated support and name the exact benchmark checkpoint/workload

Closes #342.

Source anchors

• examples/CMakeLists.txt: target server has OUTPUT_NAME vllm-server
• include/vllm.h: VLLM_ABI_VERSION 17 and the v11-v17 additions
• src/vllm/entrypoints/openai/api_server.cpp: conditional embeddings/transcription route registration
• docs/FEATURES.md plus scripts/check-supported-models.py: 35 registered architectures and the gated distinctions
• docs/BENCHMARKS.md: passing unsloth@890bdef7 grid vs speed-pending ModelOpt/current 35B grids

Verification

Passed on immutable head e91fb3fa372bbf38c645252af7f57b83bd77b4da:

• 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
• python3 scripts/check-doc-checkpoint.py --base 5812b8b6 --head HEAD
• git diff 5812b8b6..HEAD --check
• focused stale-claim scan, including scratch mutation and byte-identical restoration
• independent fresh review: PASS after one badge-scoping finding was fixed and re-reviewed

The box has no CMake executable, so the CPU configure/build and vllm-server --help execution from the spike could not run; no installation or download was attempted. The binary name is instead verified directly from CMake's OUTPUT_NAME and its registered test_server_help command. Repository-wide readiness also retains two unchanged baseline failures: CMake-dependent check-test-registration, and unrelated audit-live-rows debt.

[localai-bot]

This branch conflicts with main because #364 and #374 removed three shared record surfaces. It is a one-time cost for every branch that was in flight, and the resolution is mechanical — nothing this PR measured or decided needs revisiting. I can't push to your fork, so here is the exact recipe:

git fetch origin main && git merge origin/main

Then, per conflicted file:

File	Resolution	Why
scripts/check-public-doc-tables.py	git checkout origin/main -- scripts/check-public-doc-tables.py	The STATUS_RATCHET["chars"] key is gone. It was a byte count of docs/STATUS.md stored in the checker, so every PR that touched STATUS had to re-pin it. Your re-pin only existed to pay for your own STATUS edit; dropping it loses nothing.
.agents/NOW.md	git checkout origin/main -- .agents/NOW.md	The per-row live-claims table is gone. The live position is derived now — run scripts/now.py. Your row's next step belongs in your row's own spec under a ## Now heading, which has one writer and cannot conflict.
.agents/coordination.md	keep both sides	New claims go in .agents/claims/CLAIM-<ID>.md, one file per claim. The legacy table still parses, so an existing row there is still valid.
.agents/roadmap_v1.md	keep both sides' rows	The issue table is now sorted by row ID. Union both sides and re-sort; do not drop anyone's row.
.agents/benchmark-record.md	union both sides	Append-only log — taking either side alone silently drops a measurement.
tests/scripts/test_check_public_doc_tables.py	git checkout origin/main -- <path>	Follows the checker above.

Anything under src/, include/ or tests/*.cpp is not part of this and is a real conflict worth reading.

Sorry for the churn — removing a shared surface makes every in-flight branch pay once. After this merge your PR stops competing with other PRs for these files entirely, which is the point.

[localai-bot]

Closing: this PR has itself drifted, and landing it would swap one wrong number for another.

Its whole value is "the public docs state the right figures", and it is 258 commits behind. Measured against current main:

claim	this PR says	actual
registered architectures	35	37 (scripts/check-supported-models.py, which is the gate)

I tried to refresh the figures in place rather than close it, and stopped because they cannot be re-derived mechanically: a raw grep finds 69 REGISTER_VLLM_MODEL calls against the checker's 37 architectures, and 54 parser headers against this PR's "36 parsers". Each number has a checker-specific definition. A half-correct refresh would put newly wrong numbers into user-facing docs while looking freshly verified — worse than the current known-stale state, because nobody would think to re-check them.

Issue #342 stays open and is larger than this PR captured. On current main, README.md says 30 registered, docs/FEATURES.md says 31 in one place and 35 in another, the header exports 35 VLLM_API functions while the README says 19 symbols, and the truth for architectures is 37.

Re-cutting against current main now, deriving each count from its own checker rather than carrying any figure forward. The structural work here — the spec and the README restructure — is good and will be reused rather than discarded.