docs(readme): refresh intro, engine/backend support, and doc index

[EdHasNoLife]

Bring the public README up to date. Docs-only; independent of the in-flight SGLang/Mooncake PRs (touches only README.md).

• Conceptual on-ramp (was missing): what the project is — a cache-policy control plane, not a new cache or the gateway — the "we decide routing; the gateway follows" principle, and a one-request end-to-end flow. Points to docs/concepts/.
• "Engines and backends" section (was absent; README read LMCache-only): honest status — vLLM + LMCache/External supported today; SGLang (MP mode) + Mooncake (mesh) in active development.
• The six v1alpha1 CRDs have landed — dropped the stale "as they land".
• Generalized the webhook description from "LMCache engine configuration" to "chosen by the matched runtime adapter".
• Refreshed the Documentation section to point at the real docs/ tree (concepts, design, reference, operations, observability, reference-stack, quickstart) instead of 3 design docs.

All links verified to resolve; no naming/leakage.

[github-actions]

Codex review

Files reviewed

Root:

• README.md [reviewed]

Findings

Blocking

None.

Should-fix

README.md:27 — Replace nonexistent spec.externalEndpoint with spec.endpoint. Following the documented example currently creates an invalid CacheBackend; the v1alpha1 API and design contract expose the external address as spec.endpoint.

README.md:28 — Mooncake is incorrectly categorized as “in active development.” The current tree implements and registers the vLLM/Mooncake adapter, reconciles its managed workload, and provides a runnable sample; the README should describe its actual support status and any remaining limitations.

Nit

None.

Per-category coverage

1. Vendor-neutral naming: Clean
2. Contract & spec / docs sync: findings above at: README.md:27, README.md:28
3. gRPC fail-open semantics: Clean
4. Quality (tests / error wrapping / package placement / generated drift): Clean

Verdict

changes-requested

[github-actions]

Codex review

Files reviewed

• Root
  • README.md [reviewed]

Findings

Blocking

None.

Should-fix

• README.md:18 — “never … prompt text” overstates the metadata-only guarantee — LookupRouteRequest.prompt_text explicitly allows raw prompt text to transit the server for tokenization; only cache-state updates and stored index entries are guaranteed never to carry or retain it.
• README.md:27 — The External backend field name is incorrect — The CRD and design contract use spec.endpoint, not the nonexistent spec.externalEndpoint, so the documented configuration will not work.
• README.md:28 — Mooncake is incorrectly labeled “in active development” — The base already ships and registers the vLLM/Mooncake runtime adapter, and the design contract lists vllm/Mooncake among shipping supported pairs; distinguish that status from the genuinely unfinished SGLang MP data plane.

Nit

None.

Per-category coverage

1. Vendor-neutral naming: Clean
2. Contract & spec / docs sync: findings above at: README.md:18, README.md:27, README.md:28
3. gRPC fail-open semantics: Clean
4. Quality (tests / error wrapping / package placement / generated drift): Clean

Verdict

changes-requested