chore(release): pseudonymize attribution and scrub private-upstream references

- LICENSE/pyproject: credit "Appy" instead of legal name, consistent with the
  pseudonymous git identity.
- Remove comments naming the private upstream's internal structure
  (engine/config/facts docstrings).
- Soften README and rule/config comments that disclosed held-back internals
  (taint-tracking loader, production-tuned weights/threshold).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Appy

.gitignore

@@ -21,5 +21,5 @@ artifact_cache/
 .coverage
 .DS_Store
 
-# Agent guidance — not shipped product, references private upstream
+# Local agent guidance — not part of the published project
 CLAUDE.md

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 Ankit Prateek
+Copyright (c) 2026 Appy
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -5,9 +5,8 @@ diffs each against its prior version, runs a **community-extensible rules engine
 escalates anything suspicious to an LLM reviewer for a verdict — so a malicious update can be caught and
 reported **before** it spreads.
 
-PyDiffWatch is the open-source sibling of a private upstream scanner. The pipeline, the rules engine, and
-a starter ruleset are all here and MIT-licensed; you bring your own compute, your own LLM endpoint, and
-your own rules.
+The pipeline, the rules engine, and a starter ruleset are all here and MIT-licensed; you bring your own
+compute, your own LLM endpoint, and your own rules.
 
 **New here?** [HOWTO.md](HOWTO.md) is a step-by-step setup-to-deployment guide (endpoints, API keys,
 cron/systemd/Docker/CI). This README is the overview.
@@ -118,14 +117,12 @@ code. Example: flag a base64 decode and an exec/eval co-occurring in the same ch
 See **[RULES.md](RULES.md)** for the full schema, the predicate reference, and worked examples — written
 for both humans and LLM assistants, so you (or your agent) can author a valid rule in a few minutes.
 
-## What's here vs. upstream
+## What's here
 
 PyDiffWatch ships the full pipeline and a baseline ruleset: import-bound primitive detection, auto-exec
 location weighting, decode/fetch/credential combos, foreign-language-source, added-dependency reputation,
-and maintainer-change signals. The private upstream additionally carries a dataflow taint-tracking
-loader detector and production-tuned weights; those are not part of this open release. The engine and
-rule format here are the same ones the community extends — and the same ones upstream publishes vetted
-rules into over time.
+and maintainer-change signals. The scoring weights and threshold are baselines you can tune. The engine
+and rule format here are the ones the community extends.
 
 ## License
 

pydiffwatch/config.py

@@ -29,7 +29,7 @@ class Config:
     db_path: Path = Path(".diffwatch/diffwatch.sqlite")
     cache_dir: Path = Path(".diffwatch/artifact_cache")
     lock_path: Path = Path(".diffwatch/diffwatch.lock")
-    # containment caps (ported verbatim from main DiffWatch)
+    # containment caps
     max_download_bytes: int = 50 * 1024 * 1024
     max_members: int = 2000
     max_member_bytes: int = 10 * 1024 * 1024
@@ -44,7 +44,7 @@ class Config:
     max_releases_per_run: int = 2000
     fetch_concurrency: int = 4
     new_package_policy: str = "surface"   # "surface" | "skip" | "full"
-    threshold_t: float = 40.0             # baseline default; production-tuned value held back in main
+    threshold_t: float = 40.0             # baseline default; tune for your tolerance
     pypi_base: str = "https://pypi.org"
     webhook_url: str | None = None
     evidence_max_chars: int = 200_000

pydiffwatch/engine.py

@@ -1,9 +1,8 @@
 """Rules-engine triage: build facts from a diff, run the loaded ruleset over them, sum weights.
 
-Drop-in replacement for main DiffWatch's hardwired `triage.triage`, emitting the same public types
-(`FiredRule`, `TriageResult`) so the orchestrator/notifier/store are unchanged. The detection opinions
+Emits `FiredRule` / `TriageResult` for the orchestrator/notifier/store. The detection opinions
 live entirely in the loaded rules (`rules/community/*.yaml`); this module owns only scoring + scope
-iteration. The decode->exec taint chain and production-tuned weights are NOT here (held back upstream)."""
+iteration."""
 from .facts import build_facts
 from .models import FiredRule, TriageResult
 from .rules import evaluate

pydiffwatch/facts.py

@@ -2,10 +2,9 @@
 primitive categories, encoded-blob signal, syntax-error, location weight; plus normalized binary/dep
 findings and the maintainer-change flag. Pure and deterministic. Static AST analysis only — no execution.
 
-The import-binding machinery (_build_import_table / _resolve_call / _PRIM_BINDINGS) is the PUBLIC half of
-main DiffWatch's triage: a call counts as a dangerous primitive only when its receiver resolves via the
-file's import table to an origin in the name's allowlist (so re.compile != exec, json.loads != decode).
-The private decode->exec dataflow taint chain is NOT part of this engine (held back upstream)."""
+The import-binding machinery (_build_import_table / _resolve_call / _PRIM_BINDINGS) is what makes primitive
+detection precise: a call counts as a dangerous primitive only when its receiver resolves via the
+file's import table to an origin in the name's allowlist (so re.compile != exec, json.loads != decode)."""
 import ast
 import math
 import posixpath

pyproject.toml

@@ -5,7 +5,7 @@ description = "Rules-driven PyPI supply-chain malware scanner (open-source PyDif
 readme = "README.md"
 requires-python = ">=3.11"
 license = { text = "MIT" }
-authors = [{ name = "Ankit Prateek" }]
+authors = [{ name = "Appy" }]
 dependencies = ["PyYAML>=6.0"]
 
 [project.optional-dependencies]

rules/community/combos.yaml

@@ -1,5 +1,5 @@
-# Co-occurring import-bound primitive categories in the same changed file. Loose combos (no dataflow):
-# the precise decode->exec taint chain ("obfuscated-loader") is NOT shipped here (held back upstream).
+# Co-occurring import-bound primitive categories in the same changed file. Loose combos (no dataflow
+# taint tracking): a decode and an exec in the same file score, without proving the bytes flow between them.
 - id: combo-fetch-exec
   description: A network fetch and an exec/eval co-occur (download-and-run a second stage).
   applies_to: code