chore: prepare 0.7.0 release with audit fixes

README.md

@@ -6,7 +6,7 @@ Context Compiler helps applications keep explicit user instructions separate fro
 
 The model writes responses. The compiler stores premise and policy rules.
 
-This package is the TypeScript implementation of the Context Compiler engine, aligned with Python 0.7 API behavior.
+This package is the TypeScript implementation of the Context Compiler engine, aligned with Python 0.7 behavior.
 
 ## What it does
 
@@ -45,12 +45,12 @@ Directive examples:
   - checkpoint restore
   - experimental preprocessor behavior
   - public API behavior
-- Core public API for engine usage and transcript compilation.
+- Core public API for engine usage and transcript replay.
 - Checkpoint APIs for saving and restoring rules plus pending clarification state.
 - Controller APIs for step envelopes, preview/dry-run, and structural state diffs.
-- Decision constants exported for host-side checks.
+- Decision constants for host-side checks.
 - Experimental preprocessor module exposed through a package subpath import.
-- Fixture parity aligned to the Python 0.7.3 fixture snapshot.
+- Fixture parity aligned with the Python 0.7.3 fixture snapshot.
 
 ## Not Included Yet
 
@@ -73,7 +73,7 @@ npm install @rlippmann/context-compiler
 ## Quick Start
 
 ```ts
-import { createEngine, get_clarify_prompt, is_clarify, is_update } from '@rlippmann/context-compiler';
+import { createEngine, get_clarify_prompt, is_clarify, is_passthrough, is_update } from '@rlippmann/context-compiler';
 
 const engine = createEngine();
 const decision = engine.step('set premise concise replies');
@@ -83,7 +83,7 @@ if (is_update(decision)) {
   console.log(engine.state);
 } else if (is_clarify(decision)) {
   console.log(get_clarify_prompt(decision));
-} else {
+} else if (is_passthrough(decision)) {
   // passthrough
 }
 ```
@@ -95,7 +95,7 @@ if (is_update(decision)) {
 - `engine.state` -> current stored premise/policy rules snapshot.
 - `engine.has_pending_clarification()` -> check whether confirmation-only input is currently required.
 - `engine.exportJson()` / `engine.importJson(payload)` -> state serialization utilities.
-- `engine.exportCheckpoint()` / `engine.importCheckpoint(payload)` -> checkpoint persistence (`authoritative_state` + pending confirmation state) for safe resume.
+- `engine.exportCheckpoint()` / `engine.importCheckpoint(payload)` -> checkpoint persistence (`authoritative_state` + pending confirmation state) that safely resumes pending confirmations.
 - `engine.exportCheckpointJson()` / `engine.importCheckpointJson(payload)` -> JSON checkpoint wrapper persistence helpers.
 - `compile_transcript(messages)` -> replay user messages and return `state` or `confirm`.
 - `engine.apply_transcript(messages)` -> replay user messages onto an existing engine instance.

demos/README.md

@@ -2,16 +2,16 @@
 
 These scripts show side-by-side LLM outcomes with and without stored compiler rules, aligned with the [Python reference demos](https://github.com/rlippmann/context-compiler/blob/main/demos/README.md).
 
-Scored demos compare three host flows:
+Scored demos compare three host-side flows:
 - baseline
-- compiler path (full transcript + injected stored rules/state)
-- compiler+compact (compacted transcript + injected stored rules/state)
+- compiler path (full transcript + injected stored rules and state)
+- compiler+compact (compacted transcript + injected stored rules and state)
 
 Demo 06 is informational (context/prompt compaction metrics), not scored.
 
 ## Result variability note
 
-LLM demo outcomes can vary across environments. In practice, PASS/FAIL patterns may differ based on:
+LLM demo outcomes can vary across environments. PASS/FAIL patterns may differ based on:
 - provider
 - client layer
 - model serving path

examples/08_controller_preview_diff.ts

@@ -2,8 +2,11 @@ import {
   DECISION_CLARIFY,
   DECISION_PASSTHROUGH,
   DECISION_UPDATE,
+  POLICY_PROHIBIT,
+  POLICY_USE,
   createEngine,
   getPolicyItems,
+  get_clarify_prompt,
   getPremiseValue,
   get_decision_state,
   is_clarify,
@@ -24,8 +27,8 @@ function summarizeState(state: EngineState): {
 } {
   return {
     premise: getPremiseValue(state),
-    usePolicies: getPolicyItems(state, 'use'),
-    prohibitPolicies: getPolicyItems(state, 'prohibit')
+    usePolicies: getPolicyItems(state, POLICY_USE),
+    prohibitPolicies: getPolicyItems(state, POLICY_PROHIBIT)
   };
 }
 
@@ -46,7 +49,7 @@ function summarizeDecision(decision: Decision): {
   if (is_clarify(decision)) {
     return {
       kind: DECISION_CLARIFY,
-      promptToUser: decision.prompt_to_user,
+      promptToUser: get_clarify_prompt(decision),
       decisionState: null
     };
   }

examples/README.md

@@ -3,7 +3,7 @@
 TypeScript examples showing how host applications keep rules and corrections consistent across turns.
 
 These examples follow the Python 0.7 behavior baseline and use only core APIs.
-Request-scoped integrations can persist compiler state with checkpoint APIs so rules and corrections stay consistent across turns.
+Request-based integrations can persist compiler state with checkpoint APIs so rules and corrections stay consistent across turns.
 
 ## 01_persistent_guardrails.ts
 
@@ -30,7 +30,7 @@ Includes single-item correction with `remove policy <item>`.
 
 ## 06_transcript_replay.ts
 
-Demonstrates transcript replay behavior with `compile_transcript(messages)` and replay onto current engine state via `engine.apply_transcript(...)`.
+Shows transcript replay with `compile_transcript(messages)` and replay onto current engine state with `engine.apply_transcript(...)`.
 
 ## 07_single_policy_correction.ts
 
@@ -39,5 +39,5 @@ Demonstrates explicit single-policy correction without `reset policies`:
 
 ## 08_controller_preview_diff.ts
 
-Demonstrates controller-layer auditability with `preview(engine, input)` and `state_diff(before, after)`.
+Shows controller-layer auditability with `preview(engine, input)` and `state_diff(before, after)`.
 Shows that preview does not mutate live engine state, then applies the same input with `step(engine, input)`.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rlippmann/context-compiler",
-  "version": "0.6.1",
+  "version": "0.7.0",
   "description": "Store AI rules and corrections separately from chat history so they stay consistent across turns.",
   "keywords": [
     "llm",