fix(antigravity): strip trailing model turn for native Gemini requests too

[diegosouzapw]

Closes #10104

Root cause

AntigravityExecutor.transformRequest() in open-sse/executors/antigravity.ts only
wired stripTrailingAntigravityAssistantTurn() into the isClaude branch. A native
Gemini model routed through Antigravity (agy/gemini-*, e.g. gemini-3-flash-agent)
that ends the conversation on a role: "model" turn keeps that trailing entry
unchanged, and newer Gemini endpoints reject that shape with HTTP 400 "Requests
ending with a model turn are not supported."

stripTrailingAntigravityAssistantTurn (already correct, guarded pop-loop, never
empties contents) is now also applied on an isGemini branch
(upstreamModel.toLowerCase().includes("gemini")), independent of the
sanitizeAntigravityGeminiRequest() Claude-only field whitelist. The Claude path is
byte-identical to before — this only widens the strip to native Gemini requests.

Regression test

tests/unit/antigravity-claude-prefill-strip.test.ts:

• Test (b) previously asserted the buggy pass-through for native Gemini
  (antigravity/gemini-3.1-pro) — flipped to assert the fixed behavior (trailing
  model turn stripped, contents ends on user).
• New test (b2) covers the current callable gemini-3-flash-agent tier from PR refactor(antigravity): align official clients and callable catalog #8013's
  catalog.

Fail→pass evidence (same test file, run against pre-fix vs post-fix
antigravity.ts):

# pre-fix (HEAD~1 antigravity.ts, same test file)
✖ (b2) native Gemini agent tier (gemini-3-flash-agent, #10104) also gets the strip
  AssertionError [ERR_ASSERTION]: trailing model turn should be stripped for native Gemini
  2 !== 1

# post-fix
✔ (a) strips a single trailing assistant (model) turn for Claude models
✔ (b) strips a trailing model turn for native Gemini models too (#10104)
✔ (b2) native Gemini agent tier (gemini-3-flash-agent, #10104) also gets the strip
✔ (c) a Claude conversation already ending on user is unchanged
✔ (d) multiple trailing model turns are all stripped
✔ (e) never strips contents down to empty
✔ empty/missing contents does not throw
7 pass, 0 fail

Gates run

• npm run typecheck:core — clean
• npx eslint --suppressions-location config/quality/eslint-suppressions.json open-sse/executors/antigravity.ts tests/unit/antigravity-claude-prefill-strip.test.ts — clean
• node scripts/check/check-file-size.mjs — no new violations on touched files
• node scripts/check/check-test-discovery.mjs — OK, test file discovered
• Targeted suite node --import tsx/esm --test tests/unit/antigravity-claude-prefill-strip.test.ts — 7/7 pass
• node scripts/check/check-complexity.mjs, check-cognitive-complexity.mjs, and the full npm run test:unit were kicked off but did not finish within this session due to extreme host contention (13 parallel worktree sessions running the same checks concurrently, load average ~209). No failures observed in the partial output before the session ended; the change is a small, additive conditional branch unlikely to move complexity/cognitive thresholds. Recommend a follow-up CI run confirms these before merge.

Notes

• Hard Rule fix(ci): add environment for npm token access #18 live-VPS validation (400→200 against a real Antigravity connection) was
  not performed — no reachable live Antigravity connection in this session. The TDD
  fail→pass proof above is the validation path per the plan file's own fallback.
• Claude path (isClaude branch, PR fix(antigravity): strip trailing assistant prefill for Vertex Claude models — relocated to live executor, NEEDS LIVE VALIDATION #6114) is untouched — byte-identical.

⚠️ base-red inherited: #9985 — ESLint errors (2) from #10250 (unrelated to this change, do not chase).

[diegosouzapw]

The focused suite passes (7/7), and the Claude path remains covered. I found one plausible correctness concern: upstreamModel.includes("gemini") now strips the trailing model turn for every Gemini-named Antigravity model, including catalogued Gemini 2.5/image IDs, while the issue evidence is specifically about newer endpoints; please narrow or justify that scope. Also, issue #10104's validation plan calls out Gemini 3.6 Flash tiers, but the added tests cover only Gemini 3.1 Pro and gemini-3-flash-agent. The PR notes that no live Antigravity 400→200 check was possible; the focused TDD evidence should remain the basis for the claim unless live validation is added.