docs: drop the Go port framing, stand on its own

README.md

@@ -12,26 +12,23 @@ wake up to a branch of accepted commits and a log of every experiment,
 including the failures.
 
 Inspired by [karpathy/autoresearch](https://github.com/karpathy/autoresearch).
-This is a Python port of [g4lb/autor3search-go](https://github.com/g4lb/autor3search-go),
-its Go sibling — same design, same guarantees, a different metric source: where
-the Go harness reads `ns/op` out of `go test -bench`, this one reads per-round
-timings out of [pytest-benchmark](https://pytest-benchmark.readthedocs.io/).
+The metric source is [pytest-benchmark](https://pytest-benchmark.readthedocs.io/):
+the harness reads per-round timings out of its JSON output.
 
 > **Status: early but working.** Validated against one real library —
 > [`humanize`](https://github.com/python-humanize/humanize) — where it found and
 > kept a genuine 5.81% win across its 15 benchmarks (hoisting a per-call
 > `import math` out of nine function bodies; `clamp` −23.2%, `apnumber` −14.9%,
-> no regressions). That is one library, not the three the Go original was
-> exercised against, so treat this as a tool that inherited a validated design
-> and has begun earning its own record rather than one that already has it.
+> no regressions). That is one library, so treat this as a tool that has begun
+> earning its record rather than one that already has it.
 >
 > Every number in this README is a real measurement taken on the machine that
 > wrote it, never an illustration. Where a number would have been guessed, there
 > is no number.
 >
 > The decision procedure — the scoring rules, the Bonferroni correction, the
-> asymmetric regression guard, the four exit codes — is carried over unchanged
-> from the Go original. If you run this against your own project, the harness's
+> asymmetric regression guard, the four exit codes — is fixed, not tuned per
+> project. If you run this against your own project, the harness's
 > `results.tsv` and `report` output are the honest record of what it actually did
 > there; that is rather the point of the whole design.
 
@@ -391,10 +388,9 @@ abandoned. Plain `stop` is unaffected and behaves identically everywhere.
 you about less. One residual race is real: a grandchild started in the
 microseconds between spawning a benchmark subprocess and putting it in its job
 object is outside that job and would survive a timeout kill. Windows offers no
-way to create a process directly into a job through `subprocess`, and the Go
-original has the same window.
+way to create a process directly into a job through `subprocess`.
 
-Ported from the Go original:
+Accepted limits of the design:
 
 - No attempt to make the harness tamper-proof against a same-user attacker.
   The worktree-integrity check catches accidental clobbering and careless

src/autor3search_python/config.py

@@ -174,7 +174,7 @@ def validate(cfg: Config) -> None:
         raise ConfigError(f"gc must be one of {list(ALLOWED_GC)}, got {cfg.gc!r}")
     if _ITERATION_COUNT_FORM.match(str(cfg.benchtime)):
         raise ConfigError(
-            f"benchtime {cfg.benchtime!r} uses go test's fixed-iteration-count form (Nx), "
+            f"benchtime {cfg.benchtime!r} uses the fixed-iteration-count form (Nx), "
             f"which is deliberately unsupported: a fixed count makes rounds incomparable, "
             f"because a candidate that is twice as fast finishes in half the wall time and "
             f"is therefore measured under different thermal conditions — exactly what the "

src/autor3search_python/runner.py

@@ -309,7 +309,7 @@ def python_run(self, *args: str) -> Result:
         return self.run(self.python, *args)
 
     def compile_gate(self, paths: Sequence[str]) -> Result:
-        """The syntax gate. `go build`'s role: is this even valid code?"""
+        """The syntax gate: is this even valid code, before anything runs it?"""
         return self.python_run("-m", "compileall", "-q", "-x", _compile_exclude(), *paths)
 
     def import_gate(self, modules: Sequence[str]) -> Result:

src/autor3search_python/stats.py

@@ -1,9 +1,9 @@
 """The statistics behind a verdict.
 
-The Go original gets these from golang.org/x/perf/benchmath. Implementing them
-here keeps the harness free of a scipy dependency, and — more usefully — keeps
-the exact small-sample behavior under this project's own tests, since small
-samples are the entire operating regime (`count` defaults to 10 per side).
+These are implemented here rather than taken from scipy: it keeps the harness
+dependency-free, and — more usefully — keeps the exact small-sample behavior
+under this project's own tests, since small samples are the entire operating
+regime (`count` defaults to 10 per side).
 """
 
 from __future__ import annotations

tests/test_config.py

@@ -130,7 +130,7 @@ def test_parse_duration_rejects(text):
 
 
 def test_benchtime_rejects_the_iteration_count_form(tmp_path):
-    """go test's `100x` form is a real flag value someone may copy across; say why not."""
+    """`100x` is a plausible thing to type for a benchtime; say why it is refused."""
     with pytest.raises(config.ConfigError, match="fixed-iteration"):
         config.load(write(tmp_path, 'benchtime = "100x"\n'))
 

tests/test_e2e.py

@@ -72,10 +72,9 @@ def test_full_run(demo_repo):
     # times the largest no-op delta observed, and well clear of the KEEP step
     # below, which turns a quadratic per-character concatenation into a single
     # join and measures -25.8% (score 0.7418) on the machine this was written
-    # on. The Go original raises the same floor for the same reason; it uses
-    # 15%, against a fixture whose true effect is an order of magnitude rather
-    # than this one's quarter, so the ratio here is the thing being matched,
-    # not the number.
+    # on. What matters is that ratio — the floor sitting well inside the gap
+    # between the two effects — not the absolute number, so a fixture with a
+    # different true effect would want a different floor.
     raised = cfg_path.read_text().replace("min_effect_pct = 1.0", "min_effect_pct = 10.0")
     assert "min_effect_pct = 10.0" in raised, "init's config no longer has the key this rewrites"
     cfg_path.write_text(raised)

tests/test_runner.py

@@ -43,8 +43,8 @@ def test_timeout_is_reported_not_raised(tmp_path):
 
 
 def test_timeout_kills_grandchildren(tmp_path):
-    """go test runs its benchmark as a grandchild; pytest does too. A survivor
-    burns CPU and corrupts every later measurement on the machine."""
+    """pytest runs its benchmark as a grandchild. A survivor burns CPU and
+    corrupts every later measurement on the machine."""
     marker = tmp_path / "alive.txt"
     # marker.as_posix(): a Windows path interpolated raw would put "\\U" of
     # C:\\Users into the inner script's own non-raw string literal, and the

tests/test_templates.py

@@ -1,7 +1,26 @@
+import pathlib
 import re
 
 from autor3search_python import pipeline, templates, verdict
 
+# Names that do not exist in this project: files it never writes, a config
+# format it does not read, metrics it does not report. Naming one in the
+# agent-facing template or the README sends the reader after something that
+# is not there.
+ABSENT_NAMES = (
+    "autor3search-go",
+    "go.mod",
+    "go.sum",
+    "_test.go",
+    "config.yaml",
+    "ns/op",
+    "allocs_delta",
+    "go test",
+)
+
+# Prose that would explain this project as a port rather than on its own terms.
+PORT_FRAMING = ("Go original", "Go sibling", "Go harness", "Python port")
+
 
 def test_program_md_exit_code_table_agrees_with_verdict():
     """A stale table would send the agent branching on the wrong exit code."""
@@ -32,20 +51,27 @@ def test_program_md_documents_every_exit_code():
         assert token in text
 
 
-def test_program_md_names_no_go_artifacts():
-    """A stale Go reference would send the agent looking for a file that is not there."""
+def test_program_md_names_nothing_absent():
+    """A name for a file that is not there sends the agent looking for it."""
     text = templates.program_md()
-    for stale in (
-        "autor3search-go",
-        "go.mod",
-        "go.sum",
-        "_test.go",
-        "config.yaml",
-        "ns/op",
-        "allocs_delta",
-        "go test",
-    ):
-        assert stale not in text
+    for absent in ABSENT_NAMES:
+        assert absent not in text
+
+
+def test_readme_names_nothing_absent():
+    """The README stands on its own; it does not explain this project as a port."""
+    text = (pathlib.Path(__file__).parent.parent / "README.md").read_text(encoding="utf-8")
+    for absent in ABSENT_NAMES + PORT_FRAMING:
+        assert absent not in text
+
+
+def test_shipped_source_names_nothing_absent():
+    """Error messages and docstrings reach users too, so they get the same guard."""
+    src = pathlib.Path(__file__).parent.parent / "src"
+    for path in sorted(src.rglob("*.py")):
+        text = path.read_text(encoding="utf-8")
+        for absent in ABSENT_NAMES + PORT_FRAMING:
+            assert absent not in text, f"{path.relative_to(src)} names {absent!r}"
 
 
 def test_program_md_forbids_editing_conftest():