fix: harden impact LLM JSON parsing

.gitignore

@@ -116,6 +116,15 @@ activemq-data/
 .cfg/
 /.cfg.toml
 
+# Local project notes and recording drafts
+docs/project-notes/
+*.mp4
+*.mov
+*.webm
+*.mkv
+*.avi
+*.m4v
+
 # Local secrets and overrides
 *.local
 .streamlit/secrets.toml

README.md

@@ -103,25 +103,39 @@ system, or schema migration tool.
 
 ## Install
 
-From a checkout:
+From PyPI:
+
+- [`cfgit`](https://pypi.org/project/cfgit/)
+- [`cfgit-impact`](https://pypi.org/project/cfgit-impact/) for optional impact narration
 
 ```bash
-python -m venv .venv
-. .venv/bin/activate
-pip install -e '.[mongo,postgres,mcp,dev]'
-pip install -e plugins/cfg_impact
+pip install 'cfgit[mongo,postgres,mcp]'
 ```
 
 Minimal install for Mongo only:
 
 ```bash
-pip install -e '.[mongo]'
+pip install 'cfgit[mongo]'
 ```
 
 Minimal install for Postgres only:
 
 ```bash
-pip install -e '.[postgres]'
+pip install 'cfgit[postgres]'
+```
+
+Optional impact plugin:
+
+```bash
+pip install cfgit-impact
+```
+
+If you use `pipx`, install cfgit first and inject the optional plugin into the
+same isolated environment:
+
+```bash
+pipx install 'cfgit[mongo,postgres,mcp]'
+pipx inject cfgit cfgit-impact
 ```
 
 ## Quick start
@@ -453,7 +467,7 @@ explains in plain language what the change does, what it ripples into, and how t
 roll it back:
 
 ```bash
-pip install -e plugins/cfg_impact
+pip install cfgit-impact
 cfg impact agent_configs:agent_planner --llm --json
 ```
 
@@ -526,10 +540,7 @@ against a safe environment.
 - [docs/AGENTS.md](docs/AGENTS.md): MCP, skill, and impact plugin usage
 - [docs/SPEC_CORE.md](docs/SPEC_CORE.md): project framing and v1 scope
 - [docs/SPEC.md](docs/SPEC.md): deeper engine reference
-- [docs/README.md](docs/README.md): full documentation index, including archived project notes and historical specs
-- [docs/project-notes/findings.md](docs/project-notes/findings.md): implementation findings from the origin build
-- [docs/project-notes/handoff.md](docs/project-notes/handoff.md): archived handoff notes
-- [docs/archive/spec-v0.1.md](docs/archive/spec-v0.1.md): original historical spec
+- [docs/README.md](docs/README.md): full documentation index
 
 ## Development
 

docs/ADAPTERS.md

@@ -8,7 +8,7 @@ details. Core owns hashing, history semantics, drift detection, and restore logi
 Install:
 
 ```bash
-pip install -e '.[mongo]'
+pip install 'cfgit[mongo]'
 ```
 
 Configuration:
@@ -45,7 +45,7 @@ Mongo notes:
 Install:
 
 ```bash
-pip install -e '.[postgres]'
+pip install 'cfgit[postgres]'
 ```
 
 Configuration:

docs/AGENTS.md

@@ -54,7 +54,7 @@ tokens through MCP unless that client/session is trusted for secrets.
 Install the MCP extra:
 
 ```bash
-pip install -e '.[mcp]'
+pip install 'cfgit[mcp]'
 ```
 
 Run:
@@ -120,7 +120,7 @@ as token/DB-principal based rather than author-string based.
 Install:
 
 ```bash
-pip install -e plugins/cfg_impact
+pip install cfgit-impact
 ```
 
 Deterministic impact:

docs/README.md

@@ -12,11 +12,3 @@ Technical references:
 
 - [Core spec](SPEC_CORE.md): framing and v1 scope.
 - [Engine spec](SPEC.md): detailed engine and adapter reference.
-
-Project notes and historical artifacts:
-
-- [Implementation findings](project-notes/findings.md): source-code and data-shape findings from the origin implementation.
-- [Review findings](project-notes/review-findings.md): implementation audit defects and fix order.
-- [Session handoff](project-notes/handoff.md): archived implementation handoff.
-- [Session handoff JSON](project-notes/handoff.json): machine-readable handoff archive.
-- [Original v0.1 spec](archive/spec-v0.1.md): early historical spec.

docs/SPEC_CORE.md

@@ -189,7 +189,7 @@ writers = ["alice@*", "bob@*", "carol@*"]
 admin_actions = ["init", "restore_system"]
 ```
 
-> **First-task gate (archived in `docs/project-notes/handoff.md`):** before building, pin each collection's real id + live-rule against the actual data. `modelgarden_models` already has duplicate `model_id` values in local dev data, while `model_path` is unique and runtime-authoritative in the code paths checked. Use `model_path` unless production data proves otherwise. Getting this wrong makes the tool wrong on day one.
+> **First-task gate for any real deployment:** before tracking a collection, pin the stable id field and live-record rule against actual data. If multiple rows or documents can share the apparent id, configure a different id field or an explicit `live_when` selector before importing history. Getting this wrong makes the tool wrong on day one.
 
 ---