release: v0.6.5 — CLI hygiene + store-aware doctor

.claude/settings.json

@@ -24,7 +24,11 @@
       "Bash(git fetch *)",
       "Bash(git merge *)",
       "Bash(git add *)",
-      "Bash(git commit *)"
+      "Bash(git commit *)",
+      "Bash(\"C:\\\\Users\\\\Eric\\\\AppData\\\\Roaming\\\\Python\\\\Python313\\\\Scripts\\\\uv.exe\" run *)",
+      "Bash(git --no-pager diff .claude/settings.json)",
+      "Bash(echo \"exit=$?\")",
+      "Bash(git pull *)"
     ]
   }
 }

CHANGELOG.md

@@ -6,6 +6,35 @@ follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [0.6.5] — 2026-06-12
+
+Patch release. CLI hygiene and a store-aware `doctor` in the
+*evidence-preservation* consumer surface, plus anti-entropy guards and
+doc-freshness tripwires across the architecture and evidence-protocol docs.
+No verdict, perturbation, invariant, or spec-language change.
+
+### Changed
+
+- **Read-only commands stay off the model stack.** `doctor`, `verify`,
+  `replay`, `inspect`, `diff`, `history`, `timeline`, `matrix`, and `export`
+  no longer import `litellm` (and its import-time warnings). `main.py`
+  dispatches each command module lazily, and `execution/__init__.py` defers
+  `LiteLLMAdapter` via PEP 562 `__getattr__` (the public import is unchanged).
+  A subprocess-per-module meta-guard
+  ([`tests/meta/test_cli_import_hygiene.py`](tests/meta/test_cli_import_hygiene.py))
+  keeps it that way and also forbids read-only commands from importing
+  `falsifyai.verdict.resolver` — mechanizing the "consumers never re-resolve"
+  guarantee and closing a gap where `replay` previously had no such check.
+  (#84, #87)
+
+- **`doctor` is store-backend aware.** It now resolves the store scheme the
+  same way `build_store` does: reports the selected scheme and registered
+  backends, **fails (exit 3)** when no backend is registered for a
+  `--store-path` scheme (a missing store plugin surfaces in diagnostics
+  instead of crashing later at `run`), and write-probes only the built-in
+  SQLite store. A registered plugin store (e.g. `postgres://…`) is reported
+  but never constructed, preserving doctor's diagnose-only contract. (#85)
+
 ### Tests
 
 - **Consumer-side verdict-map coverage guard.** A new meta-test
@@ -57,6 +86,13 @@ follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   re-resolve" guarantee and closing a gap where `replay` had no such check. No
   runtime behavior changes.
 
+- **Stale phase/MVP prose refreshed** with doc-surface drift guards that fail
+  if the documented surface drifts from the code. (#93, #94)
+
+- **Agent-context correction.** `.claude/CLAUDE.md` and `AGENTS.md` no longer
+  claim the subpackages are unimplemented stubs — the pipeline runs end-to-end;
+  both now point at the CHANGELOG instead. (#86)
+
 ## [0.6.4] — 2026-06-05
 
 Patch release. Closes a self-falsification in the *evidence-generation* layer

README.md

@@ -18,7 +18,7 @@ falsifyai diff baseline candidate
 [![Python](https://img.shields.io/badge/python-3.13%2B-blue)](https://www.python.org)
 [![License](https://img.shields.io/badge/license-Apache%202.0-green)](LICENSE)
 
-**Status:** 0.6.4 — Perturbation-validity integrity. The `paraphrase` validity gate now rejects *lossy* rewrites — an `llm_rewrite` that drops a task's grounding while keeping its vocabulary — via a bidirectional-NLI entailment check under `--nli` ([case study 06](docs/case-studies/06-perturbation-validity-omission.md)); previously such an invalid perturbation could manufacture a false `CONSISTENTLY_WRONG`. `--nli`-less runs are byte-identical; spec language and replay format stay backward-compatible across the 0.x line.
+**Status:** 0.6.5 — CLI hygiene and a store-aware `doctor`. Read-only commands (`doctor`, `verify`, `replay`, `inspect`, `diff`, `history`, `timeline`, `matrix`, `export`) no longer load the model stack, and `doctor` now resolves the store backend your `--store-path` selects — failing fast when no backend is registered instead of crashing later at `run`. Atop perturbation-validity integrity (the `paraphrase` gate rejects *lossy* rewrites under `--nli`, [case study 06](docs/case-studies/06-perturbation-validity-omission.md)). Spec language and replay format stay backward-compatible across the 0.x line.
 
 ```bash
 pip install falsifyai
@@ -319,7 +319,12 @@ Consumer surfaces (`replay`, `inspect`, `diff`, `history`, `verify`, `export`) e
 
 ## Status and roadmap
 
-**0.6.4 (current release) — Perturbation-validity integrity.** Closes a generation-layer self-falsification ([case study 06](docs/case-studies/06-perturbation-validity-omission.md)): the `paraphrase` validity gate used embedding cosine, which preserves topic but not task completeness, so a rewrite that deleted a task's grounding passed and manufactured a false `CONSISTENTLY_WRONG`:
+**0.6.5 (current release) — CLI hygiene + store-aware `doctor`.** Hardens the *evidence-preservation* consumer surface without touching verdict semantics:
+
+- ✅ **Read-only commands stay off the model stack** — `doctor`, `verify`, `replay`, `inspect`, `diff`, `history`, `timeline`, `matrix`, and `export` no longer import `litellm` (lazy command dispatch + PEP 562 deferral of `LiteLLMAdapter`); a subprocess-per-module meta-guard keeps it that way and also forbids read-only commands from re-importing the verdict resolver.
+- ✅ **Store-aware `doctor`** — reports the store scheme `--store-path` selects and the registered backends, fails (exit 3) when no backend is registered, and write-probes only the built-in SQLite store (plugin stores are reported, never constructed — preserving the diagnose-only contract).
+
+**0.6.4 — Perturbation-validity integrity.** Closes a generation-layer self-falsification ([case study 06](docs/case-studies/06-perturbation-validity-omission.md)): the `paraphrase` validity gate used embedding cosine, which preserves topic but not task completeness, so a rewrite that deleted a task's grounding passed and manufactured a false `CONSISTENTLY_WRONG`:
 
 - ✅ **Bidirectional-NLI paraphrase validity** — a paraphrase must entail the original *and* be entailed by it; an omission breaks the reverse direction and is rejected (under `--nli`, reusing the oracle NLI backend). Generation-layer only; resolver byte-identical; `--nli`-less runs unchanged.
 

falsifyai/__init__.py

@@ -1,3 +1,3 @@
 """FalsifyAI — falsification-first reliability testing for AI systems."""
 
-__version__ = "0.6.4"
+__version__ = "0.6.5"

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "falsifyai"
-version = "0.6.4"
+version = "0.6.5"
 description = "Falsification-first reliability testing for AI systems: perturb inputs, preserve replayable evidence, diff reliability across model changes."
 readme = "README.md"
 license = "Apache-2.0"

tests/unit/test_version.py

@@ -11,8 +11,8 @@
 import falsifyai
 
 
-def test_dunder_version_is_0_6_4() -> None:
-    assert falsifyai.__version__ == "0.6.4"
+def test_dunder_version_is_0_6_5() -> None:
+    assert falsifyai.__version__ == "0.6.5"
 
 
 def test_dunder_version_matches_installed_metadata() -> None: