fix(website): correct LiteLLM transcription route and guards

[LauraGPT]

Summary

• Correct both ecosystem pages to use LiteLLM openai/sensevoice with an explicit self-hosted api_base: http://127.0.0.1:8000/v1. Explain the server model alias and container reachability; mark the upstream docs PR as pending, not merged.
• Replace the incorrect website guard that required custom_openai. Require visible configuration, reject unsupported provider/checkpoint examples and malformed model/base suffixes, and match the already-published 37K+ copy.
• Update only the two affected legacy manifest hashes. Refresh the historical growth ledger so old green checks are not presented as current deployment evidence.

custom_openai is not supported for audio transcription in the verified LiteLLM path. The packaged FunASR server accepts the sensevoice service alias, not FunAudioLLM/SenseVoiceSmall as its request model. The native SDK/transport/ASGI evidence is documented in LiteLLM docs #610; ASGI inference was stubbed, not acoustic-tested.

Validation

• Added 27 regression cases: 20 routing/visibility/source cases failed before the fix; six malformed-base cases failed before the boundary fix; the real-build regression caught the stale 36K+ guard before correction.
• python -m pytest tests/test_check_funasr_website_static.py web-pages/product-site/tests/test_legacy.py -q: 70 passed (54 website checks + 16 snapshot checks).
• python -m pytest web-pages/product-site/tests -q -n 4 --dist=load: 519 passed, with JUnit confirming all 519 collected cases and zero skipped/failing/error cases. Pinned pytest-xdist 3.8.0 was installed only in the isolated validation environment; repository dependencies and tests were not changed to enable parallelism.
• Two complete site builds each validated 206 pages; diff -qr confirmed identical output.
• Production-output Chromium checks passed for Chinese/English at 1440x1000 and 390x844: HTTP 200, no page errors, no horizontal page overflow, and no description text escaping the LiteLLM card.
• Black 24.4.0 with the repository's --line-length=100, pip check, and git diff --check passed. Whole-file formatting follows the existing pre-commit configuration.

The initial 240-second and 480-second serial site-test attempts timed out and are not counted as successful runs. test_output.py rebuilds the entire site for each case; four isolated pytest workers completed the unchanged full suite in 187 seconds. The initial stale-manifest failure was corrected before the successful full run.

Boundaries

This PR changes site source and guards only. It does not publish or restart the production website, download or run speech models, boot a live LiteLLM Proxy, or approve external CI/deployments. Public pages still require a separately verified release after merge. No model/GPU or full application acceptance is claimed.

Merge Verification

Merged as 12e417f4fa4490296ea2d81c5352ea2de8e4cac2 through the ordinary SHA-guarded GitHub merge endpoint. Parents are current main 15c19598110268237a3e0ca9cdcd33c83a0b3f89 and unchanged PR head f950854db3379e80a986e779da4e7208f54c0adb; GitHub verifies the merge signature. No protection setting, review, or workflow approval was changed.

The actual merged tree 24a836e751ce1faef214aa7ec49932ad75ed1d22 exactly matches the locally tested combined tree. GitHub's earlier synthetic merge reference still pointed at the old base, so it was rejected as validation evidence. A signed local-only validation commit was used instead; the PR history was not rewritten.

• Fresh combined-tree website validation: 519 passed, plus 70 passed in the focused website/snapshot selection. JUnit confirms zero skipped/failing/error cases. The focused selection overlaps the full suite; these are not 589 unique tests.
• Fresh build produced 117 product pages and validated 206 total pages.
• Fresh complete CPU compatibility selection: 38 passed under NumPy 1.26.4 and 38 passed under NumPy 2.4.0. Adjacent runtime coverage: 24 passed, 8 subtests passed. The three SenseVoice source/test/workflow blobs are unchanged from merged fix(sensevoice): preserve SentencePiece word boundaries #3737.
• An initial combined-tree test launch failed before collection because an inherited environment flag disabled pytest plugin loading. Restoring the site's native plugin environment allowed the unchanged full suite to run; the failed launch is not counted as a pass.

All four PR-head hosted checks succeeded before merge. Post-merge workflow status is recorded separately below; premerge and local successes are not claims that the main-push workflows or production deployment have completed.

At 2026-09-29 06:44 UTC, exact merge-SHA roadmap validation and API documentation succeeded. Product site validation was still in progress: legacy document links passed, build-and-validate was running, and the dependent browser job was not yet reported. This is a timestamped snapshot, not an all-green postmerge claim.

Follow-up on 2026-09-29: Product site run 36532592558 has now completed successfully on the exact merge SHA. All three jobs succeeded: legacy-docs-links (109289386077), build-and-validate (109289386363), and browser (109293383476). The earlier in-progress snapshot is historical. Production deployment remains separate and was not performed.