docs: align speech generation surfaces

[localai-org-maint-bot]

What changed

• add the landed speech API to README News and feature coverage
• update the registered-architecture count and public C ABI surface from source
• replace stale IndexTTS-2.5 refusal text with the current render and serving behavior
• preserve the active caveats: IndexTTS ignores the reference voice, and MiniMax-Music3 has no observed composed completion

Why

The user-facing docs still said /v1/audio/speech and the synthesis ABI did not exist after both landed. The README also reported the previous registry and ABI counts.

Verification

CPU-only, no GPU or model download used:

python3 scripts/check-readme-structure.py
python3 tests/scripts/test_check_readme_structure.py
python3 scripts/check-supported-models.py
python3 tests/scripts/test_check_supported_models.py
python3 scripts/check-public-doc-tables.py
python3 tests/scripts/test_check_public_doc_tables.py
python3 scripts/check-doc-checkpoint.py --staged
python3 tests/scripts/test_doc_checkpoint.py
python3 scripts/check-agent-record.py
python3 tests/scripts/test_agent_record.py
python3 scripts/check-model-checklist.py
python3 tests/scripts/test_check_model_checklist.py
python3 scripts/check-env-doc.py
python3 tests/scripts/test_check_env_doc.py
python3 scripts/check-commit-trailers.py --range upstream/main..HEAD
python3 scripts/check-role-discipline.py --base upstream/main --head HEAD --pending-pr-head $(git rev-parse HEAD)
git diff --check upstream/main...HEAD

All passed.

[localai-bot]

Closing as stale, with the true part carried forward rather than dropped.

Superseded. The count edits here are a generation behind. This branch writes
38 registered architectures and 36 tool-parser families; the registry has 40
and docs/STATUS.md records 38 families / 42 names. 9143196c7 (#1063) landed
the correct FEATURES figures. The ABI line here says v20, 45 functions where
include/vllm.h:273 says VLLM_ABI_VERSION 21 with 46 exports.

Would revert a landed result. The docs/FEATURES.md hunk asserting
"MiniMax-Music3 has not completed a composed request yet" was true when written
and is not now — docs/FEATURES.md:211 on main records an HTTP request
observed end to end over a real socket against a music-only server (#852, #672),
and docs/USAGE.md records the returned 44100 Hz stereo WAV. Merging this would
have silently withdrawn that.

Still true, and now tracked. The docs/USAGE.md repair in this branch is
correct and still needed: main at 281e6a120 continues to tell readers "There
is no /v1/audio/speech. Text to speech is not servable" (:1646, repeated
at :3623) while src/vllm/entrypoints/openai/api_server.cpp:1116 registers
that route and include/vllm.h exports vllm_synthesize. Filed as #1112,
written from today's tree, together with the docs/FEATURES.md self-contradiction
(line 210 says IndexTTS-2.5 renders, a later line says both families refuse) and
the v20/v21 disagreement.

One process note, because it is reusable rather than about this branch: the
verification here records python3 scripts/check-doc-checkpoint.py --staged as
passing. --staged reads git diff --cached --name-only, so on a clean tree it
examines zero paths and prints OK. CI ran the same checker on the same commit
and got ERROR: commit ffc9f11b: changed README.md without touching a landing source. A green over an empty set is not evidence.

Thanks — the speech-surface direction was right, and #1112 keeps it.