Separate frozen HTTP preparation from attempts

docs/internal/llm.md

@@ -29,18 +29,23 @@ for the generation boundary and is not mapped into the provider payload.
 Construction trims and validates a nonempty configured base URL and clones any
 supplied `http.Client` so Promptkit can apply its timeout default without
 mutating the caller's client. An empty configured base remains valid because a
-resolved request target may supply the endpoint. Generation then:
+resolved request target may supply the endpoint. Generation preparation then:
 
 1. validates shared execution-setting invariants and the final selected base
    endpoint;
 2. maps the internal request into the OpenAI-compatible chat payload;
 3. validates and merges extra parameters;
 4. composes `/chat/completions` through parsed URL path operations;
-5. resolves authentication;
-6. performs the outbound request under the applicable deadlines; and
-7. decodes one strictly framed, size-bounded successful response object and
-   maps its first choice and token usage, or decodes bounded structured
-   non-success detail.
+5. resolves authentication and freezes the URL, body, and optional credential.
+
+Each execution of that prepared request creates a fresh HTTP request, body
+reader, and attempt timeout context. It closes the response body and returns
+the HTTP status, decoded response or error, reported token values and presence,
+and any valid provider `Retry-After` not-before time. A strictly decoded
+successful envelope can retain usage even if its first choice or content is
+missing. Invalid framing and non-success error bodies supply no usage. The
+ordinary `Generate` method prepares and executes once and returns no partial
+response on failure.
 
 `internal/llm` owns the set of reserved OpenAI-compatible request fields used
 when validating extra parameters. Backend registration consumes the same rule

docs/roadmap/implementation.md

@@ -252,6 +252,8 @@ No observation API yet. Validate affected packages and commit.
 
 ## Stage 3 — Separate Frozen HTTP Preparation From One Attempt
 
+**Completed:** 2026-09-23.
+
 Refactor `internal/llm/OpenAICompatibleClient` without changing its one-shot
 `Generate` behavior. Add an internal `PrepareGeneration` operation that validates
 and encodes one request, resolves credentials once, and returns private frozen