feat(models)!: remove per-model gateway images — one shared gateway, models picked at runtime (#294)

[elronbandel]

Implementation half of #294 (doctrine half: #295). The model axis now has exactly one framework image kind: the shared generic gateway. EVAL_MODEL=<provider>/<model> is the only way to pick a model; EVAL_GATEWAY_IMAGE picks only the proxy flavor (bifrost default; litellm, portkey). models/replay stays as the sole in-tree pinned exception (models rule 17).

Removed

• containers/models/{gpt-5, gpt-5.4, gpt-4.1-mini, claude-opus-4, claude-sonnet-4} — the five pinned per-model images.
• containers/core/litellm — their only base, orphaned once they are gone (gateways/litellm is independent: FROM python:3.12-slim, its own eval_logger.py).
• The litellmVersion chart knob + EVAL_LITELLM_VERSION plumbing (runtime override unimplemented since Delete dead eval-entrypoint.sh #50) and the model_missing_litellm_version_* rego gates, which matched exactly the five deleted Dockerfiles.

Retargeted guards

• tests/static/model-paths.sweep.sh — asserts the single path: default flavor + flavor swap, with the EVAL_MODEL handle always reaching the gateway at runtime.
• tests/static/check.rs::model_axis_generic_default_no_silent_model — new structural invariant: every containers/models/<name>/Dockerfile must be FROM a gateways/<flavor> image (replay excepted), so a per-model image cannot sneak back into the tree.

Custom images stay possible — downstream (models rule 1a)

docs/guides/add-a-model.md gains a "Custom gateway config (advanced, non-default)" section: teams that need a shared custom config (fixed endpoint, cost rates, alias table) mount their template on the published gateway image, or bake a template-only derivation FROM gateways/<flavor> in their own registry and select it via the untouched seams (EVAL_REGISTRY+EVAL_GATEWAY_IMAGE, or gatewayImageRef). This is the migration path for anyone who relied on a deleted pinned image: the old configs are in git history and rebuild 1:1 downstream. Already-published GHCR tags remain pullable for reproducing old runs; they just stop being rebuilt.

Fixed in passing

deploy/oc conflated the two meanings of "model" (--model bifrost, --set gatewayImage=$MODEL) and its handle plumbing (evalModel) was never read by the chart — so the gateway's EVAL_MODEL got the literal image name. Now --model is the <provider>/<model> handle (what the chart's model value actually is), --gateway is the flavor, and result paths key on the clean model label (last handle segment), matching the chart's eval.modelLabel.

Docs, PR/issue templates, and .env.example collapse to the single-path story; stale core/litellm pointers now name gateways/litellm.

Verification

• cargo test -p eval-containers (29 tests, incl. bake-target conformance) — green
• cargo test --test check (16) + --test dockerfile_inspection (6) — green
• conftest verify on the dockerfile policies — 80/80
• tests/static/model-paths.sweep.sh — green
• tests/static/helm.sweep.sh — 102 benchmarks rendered + validated, 0 failed
• cargo check --workspace --tests — clean

Rules checked against

models/RULES.md 1, 1a, 12–17 · gateways/RULES.md 16–17 · RULES.md 8–9, 15 · src/RULES.md 10, 12 · verification/RULES.md (static gates above) · contributing/RULES.md 1–3 (code-only; resolves #294 with #295).

BREAKING CHANGE: the five per-model images are no longer published; set EVAL_MODEL=<provider>/<model> on the shared gateway instead — or, if you depended on a pinned image's custom config, rebuild it downstream per the new "Custom gateway config" guide section.

Closes #294 (together with #295).