fix: repair mojibake corruption; restore RUF001/002/003 tripwire

examples/01_quickstart.py

@@ -1,4 +1,4 @@
-"""memorywire quickstart — end-to-end demo using the :class:`memorywire.Memory` facade.
+"""memorywire quickstart — end-to-end demo using the :class:`memorywire.Memory` facade.
 
 This script ingests 50 short facts, runs a few recalls, deletes one user's
 records by filter, and prints final aggregate stats. It exists primarily
@@ -9,8 +9,8 @@
 --------------
 The default :class:`memorywire.store.sqlite_vec.SqliteVecStore` lazy-loads
 ``sentence-transformers/all-MiniLM-L6-v2`` on first embed call. To keep
-this example runnable anywhere — CI, a fresh laptop, a Docker image
-without ML wheels — we inject a tiny deterministic fake embedder
+this example runnable anywhere — CI, a fresh laptop, a Docker image
+without ML wheels — we inject a tiny deterministic fake embedder
 (sha256-derived 384-d vectors). The embedder is *not* representative of
 real recall quality; it exists to make the storage layer exercise its
 ANN path without pulling sentence-transformers.
@@ -32,7 +32,7 @@
 from memorywire.store.sqlite_vec import SqliteVecStore
 
 # ---------------------------------------------------------------------------
-# Fake embedder — sha256-derived 384-d deterministic vector
+# Fake embedder — sha256-derived 384-d deterministic vector
 # ---------------------------------------------------------------------------
 
 
@@ -51,7 +51,7 @@ def fake_embedder(text: str) -> list[float]:
 
 
 # ---------------------------------------------------------------------------
-# Seed data — 50 small facts
+# Seed data — 50 small facts
 # ---------------------------------------------------------------------------
 
 # Each entry is (content, user_id) so we can demonstrate per-user filtering.
@@ -129,7 +129,7 @@ def _section(title: str) -> None:
 async def main() -> None:
     """Run the full quickstart end-to-end."""
     # Use a temp file path so the demo also exercises the on-disk path.
-    # An in-memory db works just as well — toggle the line below if needed.
+    # An in-memory db works just as well — toggle the line below if needed.
     tmp_dir = tempfile.mkdtemp(prefix="amp-quickstart-")
     db_path = os.path.join(tmp_dir, "quickstart.db")
 

examples/03_procedural_fsm.py

@@ -2,10 +2,10 @@
 
 Runnable demo of the Phase-5 :mod:`memorywire.procedural` backend:
 
-1. Build the canonical ``book-flight`` procedure from spec §7.
+1. Build the canonical ``book-flight`` procedure from spec §7.
 2. Statically validate the procedure.
 3. Drive it through the happy path:
-   ``found_options → picked → paid → receipt``.
+   ``found_options → picked → paid → receipt``.
 4. Demonstrate the ``"source": "*"`` wildcard idiom by ``cancel`` from a
    mid-flow state.
 5. Roundtrip through ``to_dict()`` / JSON / ``from_dict()`` and assert
@@ -24,7 +24,7 @@
 
 
 def build_book_flight() -> Procedure:
-    """Construct the spec §7 ``book-flight`` procedure."""
+    """Construct the spec §7 ``book-flight`` procedure."""
     return Procedure(
         name="book-flight",
         states=[

pyproject.toml

@@ -117,12 +117,7 @@ select = [
   "SIM",           # flake8-simplify
   "RUF",           # ruff-specific
 ]
-ignore = [
-  "E501",             # line length enforced by formatter
-  # Intentional typography in prose docstrings/comments/strings
-  # (em dashes, curly quotes, arrows) — deliberate, not defects.
-  "RUF001", "RUF002", "RUF003",
-]
+ignore = ["E501"]  # line length enforced by formatter
 
 [tool.ruff.format]
 quote-style = "double"

scripts/bump_version.py

@@ -7,7 +7,7 @@
 manifest version.
 
 Hatch-vcs derives the actual installed version from the latest ``v*`` git tag,
-so the in-file values are *advisory* — they exist so editors and humans can see
+so the in-file values are *advisory* — they exist so editors and humans can see
 the intended version without running ``git describe``. release-please rewrites
 all three on merge.
 

scripts/extract_abstract.py

@@ -28,7 +28,7 @@
 clean = re.sub(r"\\ref\{[^}]+\}", "", clean)
 clean = clean.replace(r"\&", "&").replace(r"\%", "%").replace(r"\#", "#").replace(r"\$", "$")
 clean = re.sub(r"\\\\", " ", clean)
-clean = clean.replace("---", "—").replace("--", "–")
+clean = clean.replace("---", "—").replace("--", "–")  # noqa: RUF001 -- deliberately emits em/en dashes
 # Strip math-mode delimiters; arXiv's text field renders math as text.
 clean = re.sub(r"\$([^$]+)\$", r"\1", clean)
 # Replace common math symbols with plain-text equivalents.
@@ -44,7 +44,7 @@
 clean = re.sub(r"\s+", " ", clean).strip()
 
 print("=" * 72)
-print(" PASTE THIS INTO arXiv's 'Abstract' field — plain text ready")
+print(" PASTE THIS INTO arXiv's 'Abstract' field — plain text ready")
 print("=" * 72)
 print()
 print(clean)

scripts/inspect_pdf.py

@@ -68,7 +68,7 @@
 for label, needle in checks:
     found = needle in all_text
     mark = "  OK " if found else " MISS"
-    # Last check is inverted — we want it NOT found
+    # Last check is inverted — we want it NOT found
     if label.startswith("Old name placeholder absent"):
         mark = "  OK " if not found else " FAIL"
     print(f"  [{mark.strip()}] {label:38s}  ({'present' if found else 'absent'})")

scripts/lib/eval_common.py

@@ -54,8 +54,8 @@ class EvalConfig:
 
     Both ``run_longmemeval.py`` and ``run_locomo.py`` instantiate this
     from argparse and pass it down through the eval loop. Keep it small
-    and JSON-serializable — the per-run JSON output embeds it for
-    reproducibility audit (paper §5 calls this out).
+    and JSON-serializable — the per-run JSON output embeds it for
+    reproducibility audit (paper §5 calls this out).
     """
 
     stores: list[str] = field(default_factory=lambda: ["sqlite-vec://./eval.db"])
@@ -73,7 +73,7 @@ class EvalConfig:
     out_csv_dir: Path | None = None
 
     def to_jsonable(self) -> dict[str, Any]:
-        """Return a JSON-serialisable dict (Path → str)."""
+        """Return a JSON-serialisable dict (Path → str)."""
         return {
             "stores": list(self.stores),
             "seeds": self.seeds,
@@ -120,7 +120,7 @@ def paired_bootstrap_ci(
     seed:
         RNG seed so the CI is reproducible across reruns.
     alpha:
-        Two-sided significance level; defaults to 0.05 → 95% CI.
+        Two-sided significance level; defaults to 0.05 → 95% CI.
 
     Returns
     -------
@@ -130,13 +130,13 @@ def paired_bootstrap_ci(
 
     Notes
     -----
-    Pure-stdlib implementation — no NumPy dependency, because the eval
+    Pure-stdlib implementation — no NumPy dependency, because the eval
     harness must run on a fresh ``pip install memorywire``
     without numpy/scipy.
 
     References
     ----------
-    Efron & Tibshirani (1993), "An Introduction to the Bootstrap" §16
+    Efron & Tibshirani (1993), "An Introduction to the Bootstrap" §16
     (paired bootstrap).
     """
     if len(a) != len(b):
@@ -269,8 +269,8 @@ class LLMGrader:
 
     Cache invalidation
     ------------------
-    Swap models → key changes → cache miss → fresh grade. Swap prompt
-    template → key changes → cache miss → fresh grade. This is
+    Swap models → key changes → cache miss → fresh grade. Swap prompt
+    template → key changes → cache miss → fresh grade. This is
     intentional: the paper's reproducibility claim hinges on the cache
     capturing the *exact* grader prompt + model used to produce the
     numbers, and on a different setup producing different numbers
@@ -358,7 +358,7 @@ def _ensure_client(self) -> Any:
     def _call_openai(self, prompt: str) -> str:
         client = self._ensure_client()
         # Backoff is exponential with jitter; we don't retry on 4xx
-        # except 429 (rate limit). 429 and 5xx → retry.
+        # except 429 (rate limit). 429 and 5xx → retry.
         last_exc: Exception | None = None
         for attempt in range(self._max_retries):
             try:
@@ -381,7 +381,7 @@ def _call_openai(self, prompt: str) -> str:
                 # Exponential backoff with jitter.
                 sleep_s = (2**attempt) + random.Random(attempt).random()
                 time.sleep(min(sleep_s, 30.0))
-        # Defensive — we only get here if max_retries == 0.
+        # Defensive — we only get here if max_retries == 0.
         raise RuntimeError(f"grader call failed after {self._max_retries} retries: {last_exc}")
 
     def grade_with_meta(
@@ -432,13 +432,13 @@ def _parse_grader_response(raw: str) -> tuple[float, str]:
 
     Tolerant on purpose: trims markdown fences, swallows trailing text,
     accepts ``score`` as either int or float, clamps to [0, 1]. Returns
-    ``(0.0, raw)`` if the response cannot be parsed at all — better to
+    ``(0.0, raw)`` if the response cannot be parsed at all — better to
     score a malformed grader reply as zero than to crash a 1000-question
     eval halfway through.
     """
     stripped = raw.strip()
     # Strip markdown fences if the grader wrapped JSON in them despite
-    # response_format=json_object — defensive, not expected.
+    # response_format=json_object — defensive, not expected.
     if stripped.startswith("```"):
         lines = stripped.splitlines()
         stripped = "\n".join(line for line in lines if not line.startswith("```"))
@@ -483,7 +483,7 @@ def __init__(self, path: Path) -> None:
         self._path = path
         # writeback=False keeps memory bounded; we use the cache as a
         # straight key-value store. The shelf is held for the grader's
-        # lifetime and closed via :meth:`close` — a ``with`` block here
+        # lifetime and closed via :meth:`close` — a ``with`` block here
         # would close it before any grader call could use it.
         self._shelf = shelve.open(str(path), writeback=False)  # noqa: SIM115
 
@@ -634,7 +634,7 @@ def stage_dataset(
 
 # Rough per-1k-token rates as of 2026-Q2. Conservative defaults; users
 # can pass their own --grader-model and we'll fall back to a generic
-# rate. We don't track *every* model — only the ones we recommend.
+# rate. We don't track *every* model — only the ones we recommend.
 GRADER_RATES_USD_PER_1K_TOKENS: dict[str, tuple[float, float]] = {
     # model -> (input_per_1k, output_per_1k)
     "gpt-4-turbo": (0.01, 0.03),
@@ -654,7 +654,7 @@ def estimate_grader_cost(
     """Cost estimate in USD for ``n_calls`` grader invocations.
 
     If the model isn't in :data:`GRADER_RATES_USD_PER_1K_TOKENS` we use
-    the ``gpt-4-turbo`` rate as a pessimistic default — better the user
+    the ``gpt-4-turbo`` rate as a pessimistic default — better the user
     overestimates and isn't surprised than the other way around.
     """
     rate_in, rate_out = GRADER_RATES_USD_PER_1K_TOKENS.get(
@@ -689,7 +689,7 @@ def per_question_store_urls(
     This helper sidesteps the issue by giving each (question, seed)
     combination its own fresh SQLite file under ``workspace``. The path
     encodes ``key`` so concurrent harness runs don't clobber each
-    other. Non-``sqlite-vec`` URLs pass through unchanged — Mem0,
+    other. Non-``sqlite-vec`` URLs pass through unchanged — Mem0,
     Letta, etc. manage their own per-tenant isolation.
 
     Parameters
@@ -707,8 +707,8 @@ def per_question_store_urls(
     -------
     ``(rewritten_urls, owned_paths)``:
 
-    * ``rewritten_urls`` — the URL list to feed to ``Memory(stores=...)``.
-    * ``owned_paths`` — the files the helper created and the caller
+    * ``rewritten_urls`` — the URL list to feed to ``Memory(stores=...)``.
+    * ``owned_paths`` — the files the helper created and the caller
       should delete after ``mem.close()``. Empty for non-sqlite-vec
       URLs.
     """
@@ -739,7 +739,7 @@ def cleanup_question_dbs(paths: Iterable[Path]) -> None:
 
     SQLite WAL/SHM journals are removed alongside the main DB so the
     workspace stays bounded across a 1000-question run. Errors are
-    swallowed — losing a stale DB file is never worth aborting the
+    swallowed — losing a stale DB file is never worth aborting the
     harness for.
     """
     for db_path in paths:
@@ -753,7 +753,7 @@ def cleanup_question_dbs(paths: Iterable[Path]) -> None:
 def build_grader_context(hits: Iterable[Any], *, max_chars: int = 4000) -> str:
     """Format a list of :class:`RecallHit` rows into a grader-facing context.
 
-    The grader doesn't see the raw recall output — it sees a flat string
+    The grader doesn't see the raw recall output — it sees a flat string
     of the top-k passages, one per line, prefixed with ``[i]``. We cap
     total length at ``max_chars`` so a runaway corpus can't blow the
     grader's context window. Truncation happens at the hit boundary

scripts/preflight_arxiv.py

@@ -1,4 +1,4 @@
-"""arXiv pre-upload preflight — runs against the checklist arXiv displays
+"""arXiv pre-upload preflight — runs against the checklist arXiv displays
 just before file upload. Catches the issues that slow down announcement:
 
   1. TeX source is present (not PDF-only).