docs: align registry and speech status

[localai-org-maint-bot]

Summary

• update the README and status ledger from 37 to 38 registered architecture names
• update the README from 36 parser families and 40 names to 37 families and 41 names
• describe the CPU-gated IndexTTS-2.5 foundations without claiming a loadable model or audio synthesis
• record the IndexTTS-2.5 benchmark disposition as pending

No News entry is added because IndexTTS-2.5 still refuses at load and is not a user-available feature.

Verification

• python3 scripts/check-readme-structure.py
• python3 scripts/check-supported-models.py
• python3 scripts/check-public-doc-tables.py
• python3 scripts/check-agent-record.py
• python3 scripts/check-model-checklist.py
• python3 scripts/check-env-doc.py
• python3 -m unittest tests.scripts.test_check_readme_structure tests.scripts.test_check_supported_models tests.scripts.test_check_public_doc_tables (107 tests)
• git diff --check

The broader test-registration gate cannot run on this host because cmake is not installed. The change is documentation-only.

[localai-bot]

Closing as stale. Every substantive claim in this branch is now false, and
merging it would silently revert accurate text.

Measured against main at 281e6a120:

• "Load refuses because the remaining stages, routes, and C ABI are absent. No
  model synthesizes audio" — the route is registered at
  src/vllm/entrypoints/openai/api_server.cpp:1116, and include/vllm.h
  exports vllm_synthesize plus the vllm_speech_* family.
  docs/FEATURES.md:210 on main describes IndexTTS-2.5 as rendering text to
  audio on real weights, with the reference clip ignored and CAMPPlus returning
  NaN (IndexTTS-2.5: port the vLLM-Omni TTS lane (GPT-2 AR talker + EnhancedCodec + S2Mel CFM/DiT + BigVGAN) — our first audio-generating model #634, vLLM-Omni has no parity pin: H3 W3+, LTX-2.5 and ~40 omni-only architectures (the whole TTS family included) cannot be gated against any oracle #633) — a narrower and different defect than "cannot load".
• "38 registered architectures" and "37 tool-parser families (41 accepted
  names)" — the registry has 40, and docs/STATUS.md records 38
  families / 42 names. inkling landed after this branch's base. 9143196c7
  (docs(FEATURES): the registry has 40 architectures, and the table said 38 #1063) landed the correct FEATURES figures.
• The docs/STATUS.md:95 hunk replaces one wrong architecture count with
  another.

The docs/BENCHMARKS.md row's PENDING disposition is defensible on its own —
#633 is unpinned and no speed number exists — but it is attached to
"no loadable model, synthesized audio ... exists", which is not the case.

What was right here, and is worth keeping: the surface discipline. A
lifecycle statement to STATUS.md, a disposition to BENCHMARKS.md, a
capability to FEATURES.md is exactly the split AGENTS.md asks for. And the
gap it noticed is real: main still has no IndexTTS-2.5 row in either
docs/STATUS.md or docs/BENCHMARKS.md, while docs/FEATURES.md describes it
as rendering. That, together with the docs/USAGE.md passages still denying the
route exists and the v20/v21 disagreement inside FEATURES, is filed as
#1112 — to be written from today's tree rather than from this one.

Also worth recording so it is not re-derived: this branch's sanitize-cpu
(address,undefined) red was never attributable. The failure was
src/vllm/model_executor/models/ltx2_loader.cpp:1325:91: runtime error: load of misaligned address ... for type 'const uint16_t' in test_ltx2_video — a C++
defect in a file this markdown-only diff does not touch. main itself carried
that red from cefacd2d0 until fc903b8dd closed it (#674), about 11 hours
after this branch's head commit. Likewise its windows-msvc-* reds: a
markdown-only diff cannot break an MSVC build, and that queue has since been
worked down through #968 (landed as #983) and #1068.