fix(ollama): set ContextSize via the embedded LLMConfig so the package builds

[localai-bot]

Problem

make lint has been failing on master since bf19758, and takes tests-linux / tests-apple down with it on every PR branched after that commit:

core/http/endpoints/ollama/helpers_internal_test.go:33:31:
  unknown field ContextSize in struct literal of type
  "github.com/mudler/LocalAI/core/config".ModelConfig
core/http/endpoints/ollama/helpers_internal_test.go:44:30: (same)

ContextSize is not a direct field of ModelConfig. It lives on LLMConfig (core/config/model_config.go:1059), which ModelConfig embeds inline (core/config/model_config.go:106). Go permits reading a promoted field but not setting one in a composite literal, so config.ModelConfig{ContextSize: &existing} has never compiled.

Master lint history pins it exactly:

commit	lint
48b7d6d8f	success
bf19758e0 (#11032)	failure
everything after	failure (inherited)

Fix

Use the literal form already used everywhere else in the tree for this field (see core/backend/options_internal_test.go:301):

cfg := &config.ModelConfig{LLMConfig: config.LLMConfig{ContextSize: &existing}}

Two lines, test-only. No production code touched.

Why this matters beyond a green build

The specs were not merely uncompiled, they were inert. #11032 is a DoS fix: an unauthenticated client raising the context ceiling drives KV-cache allocation, and an oversized num_ctx can trigger a catastrophic OOM. The regression guard it shipped was never executing.

Confirmed the restored specs are functional rather than vacuous by stubbing out the ceiling clamp in applyOllamaOptions, which fails the intended spec:

[FAIL] applyOllamaOptions num_ctx clamping (issue #11022)
  [It] does not let an oversized num_ctx raise an existing context ceiling
  helpers_internal_test.go:37

The coverage gate corroborates it: coverage rose 54.2% -> 55.3% from this two-line change alone, because a whole package's tests started running.

Verification

• go vet ./core/http/endpoints/ollama/ - clean
• go test -count=1 ./core/http/endpoints/ollama/ - ok, 5/5 num_ctx specs execute
• make lint (full repo) - 0 issues
• Full pre-commit gate incl. coverage - passed, no --no-verify

Note on process

bf19758 merged carrying a self-contained compile error in its own new test file, so its PR either never ran lint or merged on a stale check. May be worth a look at why the gate allowed it.

Unrelated CI noise (no action needed here)

While diagnosing I checked the other red PRs:

• Yamllint reds on fix(worker): never reuse a backend process whose directory a reinstall replaced #11029, chore: ⬆️ Update vllm-metal (darwin) to v0.3.0.dev20260722081849 #11034, chore: ⬆️ Update CrispStrobe/CrispASR to 3ab5f4ac13685966b47cc75dc7fd02f3c4a51beb #11035, chore: ⬆️ Update localai-org/ced.cpp to db5aae02973a745722d6fbd2157cab1999106777 #11037, chore: ⬆️ Update leejet/stable-diffusion.cpp to 8a51eb92848c1327a5aaeff5ad81a7a9a2435255 #11038, chore: ⬆️ Update ServeurpersoCom/qwentts.cpp to 82cd05b9f3a175612dc89fd6943e610fab096ef5 #11039, chore: ⬆️ Update ServeurpersoCom/omnivoice.cpp to 4f33af825d66e6ef1cb185e87b4589cacf747291 #11040 are stale. Those branches sit on 5c96e097b, before the gallery/index.yaml indentation fix. Running yamllint 1.33 against current master's gallery/ gives zero errors. They need a re-run/rebase, not a code change.
• tests-transformers on ci(backend): refuse to move mutable image tags backwards #11025 (object Result can't be used in 'await' expression) is from Jul 21 22:12 and the extras workflow is green on master since. Likely a transient unpinned-dep issue; worth a re-run before treating it as real.