Fix Gemini ACP model IDs with Google provider prefixes

[kywch]

Fixes #975

Summary

• Normalize models.dev google/gemini-* and google/gemma-* IDs to bare model IDs on the Gemini ACP path.
• Normalize the same Google model families to a single gemini/ provider prefix when routing through the generic LiteLLM gateway.
• Preserve unrelated google/* model IDs.
• Port fix(gemini): strip Google model prefix #979 stronger connect_acp -> session/set_model regression coverage for Gemini and Gemma, plus focused gateway boundary coverage.

Validation

• uv run pytest -q tests/test_acp.py tests/test_litellm_config.py tests/test_litellm_hardening.py -> 175 passed after merging current main.
• Ruff lint and format checks passed for all changed files.
• Live A/B evidence already documented in this review comment: real @google/gemini-cli@0.42.0 and Google GenerateContent returned 200 for normalized Gemini and Gemma IDs; prefixed IDs reproduced the provider 404.

[tulerfeng]

I have Gemini API credit available, so I was able to run the live checks that
this PR and #979 were both missing — a real @google/gemini-cli@0.42.0 talking
to the real Google GenerateContent endpoint, rather than only the offline
string-level reproduction.

Short version: this PR fixes the reported bug, and it fixes strictly more than
#979 does.

Setup

• merge-base 29c64e7, fix(gemini): strip Google model prefix #979 head bc9a1c3, this PR head 0845235
• @google/gemini-cli@0.42.0 (the registry pin) in a node:22-slim container
• Live Google GenerateContent API, personal API key
• The ACP modelId under test is produced by the real code path, using the
  source-level reproduction from Gemini ACP passes google/ provider prefixes into Gemini CLI model IDs #975 (_resolve_acp_model_input ->
  _format_acp_model), then handed verbatim to the CLI via -m.

Live A/B

requested model	tree	ACP modelId sent	live gemini-cli 0.42.0
google/gemini-3.1-flash-lite-preview	29c64e7 (before fix)	google/gemini-3.1-flash-lite-preview	404 ModelNotFoundError
google/gemini-3.1-flash-lite-preview	bc9a1c3 (#979)	gemini-3.1-flash-lite-preview	200 ok
google/gemini-3.1-flash-lite-preview	0845235 (this PR)	gemini-3.1-flash-lite-preview	200 ok
google/gemma-4-31b-it	29c64e7 (before fix)	google/gemma-4-31b-it	404 ModelNotFoundError
google/gemma-4-31b-it	bc9a1c3 (#979)	google/gemma-4-31b-it	404 ModelNotFoundError
google/gemma-4-31b-it	0845235 (this PR)	gemma-4-31b-it	200 ok
gemini-3.5-flash-lite (already bare)	all three trees	gemini-3.5-flash-lite	200 ok

The gemma rows are the reason I'd land this one rather than #979. #979 matches
google/gemini-, so google/gemma-* still reaches the CLI prefixed and still
404s. This PR matches google/ and normalizes both. That shape is not
hypothetical for BenchFlow: _MODELSDEV_PROVIDER_HEURISTICS already carries
("gemma", "google"), so the codebase produces google/gemma-* in the reverse
direction.

The 404 in #975, confirmed end to end

The empty error message in the original report ("message":"b''") is not a
truncation. Google returns a routing 404 with no body at all when the / in the
model name breaks the URL path:

POST /v1beta/models/gemini-3.5-flash-lite:generateContent         -> 200
POST /v1beta/models/google/gemini-3.5-flash-lite:generateContent  -> HTTP/2 404
                                                content-type: text/html
                                                content-length: 0

Which is exactly what the CLI then surfaces:

ModelNotFoundError: {"error":{"message":"","code":404,"status":"Not Found"}}

Empty body -> empty message -> b'' once it reaches BenchFlow. The causal chain
in #975 holds all the way to the provider.

Tests, lint, and mutation

• uv run python -m pytest tests/test_acp.py -q -> 96 passed
• uv run python -m pytest tests/test_litellm_hardening.py -q -> 52 passed
• uv run ruff check / ruff format --check on both touched files -> clean
• Mutation: deleting the two added lines makes
  test_gemini_acp_model_strips_google_prefix fail and restoring them makes it
  pass again, so the new test genuinely guards the fix.
• Cross-check: applying fix(gemini): strip Google model prefix #979's tests/test_acp.py diff on top of this PR gives
  14 passed in TestConnectAcpModelSelection. This implementation satisfies
  fix(gemini): strip Google model prefix #979's assertions as well; the reverse is not true because of the gemma gap.

Non-regression, identical on all three trees: opencode + gemini-3.5-flash-lite
still resolves to google/gemini-3.5-flash-lite, and pi-acp + benchflow-x
still resolves to litellm/benchflow-x.

Suggestion

Land this implementation, and consider folding in #979's parametrized case in
tests/test_acp.py — it exercises connect_acp and asserts on the actual
session/set_model call, which is a stronger guard than the direct
_format_acp_model unit test here. Adding a google/gemma-* case would pin the
behaviour that distinguishes the two approaches.

Minor: this PR describes #975 but does not use a closing keyword, so it is not
linked to the issue. #979 has Fixes #975.

Everything above was validated end to end against the live Gemini API — real
CLI, real provider, real 404 — rather than at the string level only.

[kywch]

#979 will close unmerged after this replacement lands. Its stronger connect_acp -> session/set_model coverage is now ported here for both Gemini and Gemma. Its unrelated judge change (bc9a1c3) will move to a separate PR.

Validation after merging current main: 175 affected tests passed; Ruff lint/format and git diff --check passed. Branch pushed normally and remains draft pending fresh CI/review.

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)

[kywch]

Addressed Devin’s acpx/gemini finding using existing ACPX normalization, added integration coverage, and removed duplicate unit coverage. Merged current main for RestrictedPython 8.3. All CI checks pass.