docs(ENG-MM-INPUT-PIPELINE): correct HTTP support claims

[localai-org-maint-bot]

Row

ENG-MM-INPUT-PIPELINE, public documentation correction only.

Before starting

• Local issue: ISSUE-LOCAL-01M24M9SWMN0WAX7PH2ZC2YWFN.
• Started from fresh upstream main at 51ded8f10 in a separate worktree.
• Inspected open pull requests and local claims; none owns this HTTP documentation correction.
• One pull request carries the committed source inventory, documentation edits, and review correction.
• Inspected the existing row and multimodal specs. No feature lifecycle changes; helper implementation readiness is not applicable.
• Source: runner.cpp:3094, server_main.cpp:1523-1576, and mm_chat_dots3note.cpp:132-174.
• The two served-forward test suites use synthetic weights and do not establish real-checkpoint token parity.

What changed

The README says the runner drops image features, although HTTP requests now reach the model.
Replace that claim with the implemented Qwen3-VL and dots3-note paths, and add a short news entry.
Document per-model image/audio limits, square raw RGB images, and PCM16 WAV audio.
Replace the misleading JPEG example with a file-based raw-RGB request using Python's standard library.

Shorten repeated execution details in the feature overview and link to the input guide.
Clarify that zero limits reject dots3-note requests but still load its supported towers.
Preserve the measured Qwen3-VL memory result and the pending real-checkpoint token gate.

Evidence

All checks below passed locally on reviewed head c561ce1e6, using CPU only:

python3 scripts/check-readme-structure.py
python3 scripts/check-supported-models.py
python3 scripts/check-surface-coverage.py
python3 scripts/check-quickstart-recipes.py
python3 scripts/check-agent-record.py
python3 tests/scripts/test_check_readme_structure.py
python3 tests/scripts/test_check_surface_coverage.py
python3 tests/scripts/test_check_supported_models.py
python3 tests/scripts/test_check_quickstart_recipes.py
git diff --check 51ded8f10..HEAD

The four mutation suites passed 108 tests: 19, 46, 11, and 32 respectively.
Offline execution of the documented Python example verified the POST endpoint, model name, MIME type, and exact image bytes.
Empty and malformed buffers refused before HTTP. Local links in all four edited documents resolved.
Independent review passed after a fresh implementer corrected the tower-skip scope and clarified the default-limit example.

Speed claims

This pull request makes no speed claim. No runtime code, benchmark result, or feature lifecycle changes.

Honest gaps

Full preflight was attempted and stopped after established environment failures; it did not pass.
This minimal Alpine shell lacks a C/C++ toolchain and GNU find -printf.
Some tests clear the environment and cannot load the temporary Python shared libraries.
The focused documentation checks above pass; no checker was changed or bypassed to produce those results.
No model inference or GPU gate was run. CI on the final push remains separate verification.
The local issue remains open until upstream integration.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Codex:GPT-6 [Codex]