refactor: implement Pipeline v2 execution model with declarative step metadata, YAML configuration, and unified engine delegation.

Author: tercel

CHANGELOG.md

@@ -5,6 +5,29 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.17.0] - 2026-04-05
+
+### Added
+
+- **Step Metadata**: Four declarative fields on `BaseStep`: `match_modules` (glob patterns for selective execution), `ignore_errors` (fault-tolerant steps), `pure` (safe for `validate()` dry-run), `timeout_ms` (per-step timeout).
+- **YAML Pipeline Configuration**: `register_step_type()`, `unregister_step_type()`, `registered_step_types()`, `build_strategy_from_config()` — configure pipeline steps via `apcore.yaml` at startup.
+- **PipelineContext fields**: `dry_run`, `version_hint`, `executed_middlewares` for pipeline-aware execution.
+- **StepTrace**: `skip_reason` field for understanding why steps were skipped ("no_match", "dry_run", "error_ignored").
+
+### Changed
+
+- **Step order**: `middleware_before` now runs BEFORE `input_validation` (was after). Middleware input transforms are now validated by the schema check.
+- **Executor delegation**: `call()`, `call_async()`, `validate()`, and `stream()` fully delegate to `PipelineEngine.run()`. Removed ~300 lines of duplicated inline step code.
+- **Renamed**: `safety_check` step → `call_chain_guard` (accurately describes call-chain depth/cycle/repeat checking).
+- **Renamed**: `BuiltinSafetyCheck` class → `BuiltinCallChainGuard`.
+
+### Fixed
+
+- Middleware input transforms were never re-validated against the schema (now validated after middleware runs).
+- `validate()` was hardcoded to 7 inline checks; now uses `dry_run=True` pipeline mode — user-added `pure=True` steps automatically participate.
+
+---
+
 ## [0.16.0] - 2026-04-05
 
 ### Added

README.md

@@ -17,7 +17,7 @@ A schema-enforced module standard for the AI-Perceivable era.
 ## Features
 
 - **Schema-driven modules** -- Define input/output contracts using Pydantic models (with automatic validation) or plain JSON Schema dicts
-- **Execution Pipeline** -- Context creation, safety checks, ACL enforcement, approval gate, validation, middleware chains, and execution with timeout support
+- **Execution Pipeline** -- Context creation, call chain guard, ACL enforcement, approval gate, middleware before, validation, execution, output validation, middleware after, and return -- with step metadata (`match_modules`, `ignore_errors`, `pure`, `timeout_ms`) and YAML pipeline configuration
 - **`@module` decorator** -- Turn plain functions into fully schema-aware modules with zero boilerplate
 - **YAML bindings** -- Register modules declaratively without modifying source code
 - **Access control (ACL)** -- Pattern-based, first-match-wins rules with wildcard support

src/apcore/__init__.py

@@ -198,6 +198,14 @@
     StrategyNotFoundError,
 )
 
+# Pipeline Configuration
+from apcore.pipeline_config import (
+    build_strategy_from_config,
+    register_step_type,
+    registered_step_types,
+    unregister_step_type,
+)
+
 # Trace Context
 from apcore.trace_context import TraceContext, TraceParent
 
@@ -544,6 +552,11 @@ def enable(module_id: str, reason: str = "Enabled via APCore client") -> dict[st
     "StepNotReplaceableError",
     "StepNameDuplicateError",
     "StrategyNotFoundError",
+    # Pipeline Configuration
+    "register_step_type",
+    "unregister_step_type",
+    "registered_step_types",
+    "build_strategy_from_config",
     # System Modules
     "register_sys_modules",
     "register_subscriber_type",