Add configurable transient retries for CLI execution

docs/cli.md

@@ -64,6 +64,7 @@ Optional flags:
 | `--top-p <float>` | Runtime top-p override. |
 | `--reasoning-effort <value>` | Runtime reasoning-effort override. |
 | `--timeout <duration>` | Runtime timeout override using Go duration syntax. |
+| `--transient-retries <integer>` | Nonnegative additional transient attempt count for this request; `0` disables additional attempts. |
 
 Deprecated aliases: `--prompt-id` for `--prompt`, and `--profile-id` for
 `--profile`.
@@ -77,6 +78,10 @@ zero-second override. The timeout layers are defined in the
 [Promptkit outbound integration contract](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.11.0/docs/integrations/openai-compatible-chat.md#timeout-and-cancellation).
 With the built-in client, each transient retry gets a fresh generation timeout;
 this flag does not limit the complete run.
+Omitting `--transient-retries` inherits the application setting or Promptkit's
+current default of two additional attempts. The flag overrides only this
+request; it does not alter the configured engine. Structural repair generations
+have separate transient retry budgets.
 
 There is no raw API-key flag. Use `--api-key-env`.
 
@@ -107,13 +112,15 @@ scriptorium render [flags]
 `--config`, `--prompt-dir`, `--profile-dir`, `--prompt-version`, `--profile`,
 `--input`, `--var`, `--out`, `--llm-base-url`,
 `--model`, `--api-key-env`, `--temperature`, `--max-tokens`, `--top-p`,
-`--reasoning-effort`, `--session-id`, `--timeout`, and `--format text|json`. Their meanings match the corresponding
+`--reasoning-effort`, `--session-id`, `--timeout`, `--transient-retries`, and `--format text|json`. Their meanings match the corresponding
 `run` flags; `--format` selects prepared-run output and otherwise uses
 `defaults.render_format`.
 
 The same deprecated aliases and numeric/timeout behavior as `run` apply.
 The same session and reasoning inheritance, replacement, and clearing behavior
 also applies.
+`render` validates the retry override but does not generate model output or
+include a retry field in prepared output.
 `render` does not accept `--schema-dir`; configure `schema_dir` through the
 configuration file. It resolves profiles and schemas as part of preparation but
 does not call an LLM.

docs/config.md

@@ -40,6 +40,7 @@ do not override a prior value. Raw API-key fields are not accepted.
 | `server.max_response_bytes` | `16777216` | Maximum encoded HTTP response bytes; `0` disables the limit. |
 | `defaults.render_format` | `text` | Default prepared-run output format: `text` or `json`. |
 | `backends` | unset | Optional mapping of custom Promptkit backend IDs to engine-scoped connection and capacity settings. |
+| `transient_retries` | Promptkit default | Optional nonnegative count of additional built-in transient generation attempts. Omitted or `null` delegates to Promptkit; `0` disables additional attempts. |
 
 The size fields must be zero or greater. The [HTTP API](api.md) defines how
 each limit is enforced and reported. `server.artifact_root` configures an HTTP
@@ -120,12 +121,13 @@ A positive Promptkit `repair_attempts` budget can add provider calls, latency,
 token use, and cost; see the tagged format reference for its permitted values
 and validation-mode requirements.
 
-Promptkit's built-in client now makes up to two additional attempts for eligible
-transient HTTP statuses by default. This can increase latency, request count,
-and cost. Scriptorium currently uses that upstream default without an
-application-level retry setting. Each initial or repair generation has its own
-transient budget; an injected client retains its own retry behavior. The tagged
-integration contract owns eligibility and timing.
+Promptkit's built-in client makes up to two additional attempts for eligible
+transient HTTP statuses by default. The application `transient_retries` setting
+changes that budget for its constructed engine; an explicit `run` request
+override takes precedence. This can increase latency, request count, and cost.
+Each initial or repair generation has its own transient budget; an injected
+client retains its own retry behavior. The tagged integration contract owns
+eligibility and timing.
 
 Promptkit's
 [OpenAI-compatible integration contract](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.11.0/docs/integrations/openai-compatible-chat.md)

docs/internal/adapters.md

@@ -33,6 +33,10 @@ described by its tagged
 and profile selection, optional file inputs, variables, and presence-aware
 execution overrides. Omitted framework settings remain zero values so Promptkit
 resolves its own defaults and definition-required inputs.
+The application retry setting enters shared engine construction, while a
+run/render flag enters only the mapped request. The server uses the same engine
+setting and retains one backend admission state. Render prepares without
+generation.
 
 `inspect prompt` maps the selected ID and optional version to
 `promptkit.Engine.InspectPrompt`, then formats Scriptorium-owned inspection

docs/operations.md

@@ -112,8 +112,11 @@ call.
 Promptkit's built-in client now allows two additional attempts for eligible
 transient provider statuses by default. Account for added calls, latency, and
 cost when planning capacity. Each repair generation has its own transient
-budget, separate from the structural repair count. The current Scriptorium
-interfaces use this upstream retry default. Existing generation timeouts apply
+budget, separate from the structural repair count. Configure the engine budget
+with [`transient_retries`](config.md); `run` and `render` accept a request
+override described in the [CLI contract](cli.md). Omission inherits the engine
+setting and then Promptkit's default, while explicit zero disables additional
+attempts. Existing generation timeouts apply
 per attempt, while the caller context covers the complete operation. Provider
 `Retry-After` timing can exceed Promptkit's ordinary backoff cap; consult the
 [outbound integration contract](https://gitea.maximumdirect.net/eric/promptkit/src/tag/v0.11.0/docs/integrations/openai-compatible-chat.md)

docs/roadmap/implementation.md

@@ -107,7 +107,7 @@ are handled; no new runtime flags or report schema are claimed in current docs.
 
 ## Stage 2: Wire Retry Configuration And CLI Overrides
 
-**Status: Pending.**
+**Status: Complete.**
 
 Selected interface: top-level application YAML `transient_retries`, and
 `--transient-retries <integer>` on `run` and `render`. This is not a model