docs: clarify model checkpoint and serving choices

[LauraGPT]

Summary

• Distinguish Fun-ASR-Nano split-engine, official native vLLM and historical community checkpoints across EN/ZH/JA/KO model-selection guides.
• Scope HTTP aliases to the example server and document MOSS as third-party offline transcription with anonymous per-recording speaker labels.
• Add rendered-anchor and real EN/ZH navigation tests without changing existing routes or headings.

Validation

• Documentation-focused tests: 40 passed, including a rerun at the signed source commit.
• Model-free documentation contracts: 282 passed; three existing CI exclusions retained (toy checkout and two browser-only tests).
• Sphinx link checks: 46 passed, zero skips.
• Product-site suite: 196 passed; guide sync check passed.
• Built 103 product pages; validated 186 HTML pages. Signed-commit rebuild matched all 248 output files byte for byte.
• Chromium journeys: 10 passed across EN/ZH 320/390/1440px; ten screenshots inspected. Four additional mobile horizontal-scroll probes passed.
• Independent source review found no blockers; reviewed hashes match the committed candidate.

Boundaries

Documentation and tests only: no new model/runtime support, acoustic quality, known-person identification, sustained-load or real-time guarantee. Historical community measurements are not relabeled as official model results. The new container initially lacked Chromium system libraries; the unchanged browser tests passed after installing the official dependencies, and the initial failure logs were retained.

Backup: model-choice-precommit-20260908T182854Z on ind-gpu8. Signed commit with DCO; publication requires successful exact-head CI and fresh base/merge-tree verification.

CI follow-up at 53fbe89

The initial vLLM contract job exposed a missing bs4 dependency introduced by placing renderer checks in its source-contract module. This was reproduced in a pytest-only environment (collection exit 2), then fixed by moving both language cases into the existing product-site build-fixture suite. No checks were removed or skipped; anchor uniqueness now covers the full rendered page. The pytest-only suite passes 18 tests, the combined suite passes 40 tests at the new signed head, and rebuilding still produces the same 248 files byte for byte. The earlier 196 product / 282 documentation counts above refer to the initial test placement; two rendering cases now belong to the product suite. Full exact-head CI remains the merge gate.