feat(model): complete Qwen3.8 support - #5339

Re-checked all four fixes in b026812d2 against the new head — all confirmed, no regressions. LGTM.

P1 MLX cache aliasing — list(cached_tokens) matches upstream's own cache_key[:] at mlx_lm/server.py:874. The test is a real one: it keeps the generator-owned list, appends to it after insertion, and asserts the stored key is a distinct object.

P2 malformed sibling call — re-ran my original repro against both revisions:

BAD then GOOD, same delta:     main: RAISED KeyError   ->  now: (None, 'good', {'x': 1}, 1)
BAD first, GOOD later delta:   main: RAISED KeyError   ->  now: (None, 'good', {'x': 1}, 1)

Now strictly better than main, which still raises in both. I also checked index bookkeeping after a skip: the malformed call keeps its slot, so with good / malformed / good the survivors emit as index 0 and 2 with stable call IDs and a single name emission each. Non-contiguous indexes are correct here — index identifies the call, and the alternative would renumber calls the client may already have seen. The partial-name path is regex-only, so it can't raise.

P2 CA bundle — verified the narrowed fix does both things it needs to:

HTTPS_PROXY + ALL_PROXY + REQUESTS_CA_BUNDLE set, trust_env=True,
proxies={"http": None, "https": None, "all": None}
  -> selected proxy: None
  -> verify: /nonexistent/ca.pem    (CA bundle preserved)

P3 docs — scheme-dependent wording is accurate now, and the code comment at model/utils.py:663 was updated to match.

All three new regression tests fail on b0f47155d and pass on b026812d2, so they're genuinely covering the fixes. Also re-ran the earlier streaming-order cases (prefix content, inter-call content, close/content/open, complete-following-call) — all still correct.

One thing I did not verify: no MLX hardware here, so the prompt-cache path is only checked against upstream mlx-lm's logic and the unit tests, not a live batch run.