Sync from source tree: verify.py first-command docs + measured test badge

README.md

@@ -21,7 +21,7 @@ configs:
 
 # sparam-conformance
 
-![CI](https://github.com/nickharris808/sparam-conformance/actions/workflows/ci.yml/badge.svg) ![Licence](https://img.shields.io/badge/data-CC--BY--4.0-green) ![Cases](https://img.shields.io/badge/cases-11%20labelled-blue) ![Tests](https://img.shields.io/badge/tests-57%20passing-brightgreen)
+![CI](https://github.com/nickharris808/sparam-conformance/actions/workflows/ci.yml/badge.svg) ![Licence](https://img.shields.io/badge/data-CC--BY--4.0-green) ![Cases](https://img.shields.io/badge/cases-11%20labelled-blue) ![Tests](https://img.shields.io/badge/tests-61%20collected-brightgreen)
 
 📖 **[Documentation site](https://nickharris808.github.io/physics-lint/)** — the portfolio narrative, the concepts, a full walkthrough, and what all of this proves (and does not).
 
@@ -43,10 +43,41 @@ opinion. We know `active_gain` is non-passive because we built 3× gain into it.
 ```bash
 git clone https://github.com/nickharris808/sparam-conformance.git && cd sparam-conformance
 
+python3 verify.py                                     # check what you downloaded
 python score.py --checker mypackage.mychecker:run     # grade your checker
 python generate.py                                    # optional: rebuild the corpus
 ```
 
+`verify.py` is the one command that works before you have anything else. It uses
+only the standard library — no install, no dependencies, nothing from the
+repository that produced the corpus:
+
+```
+$ python3 verify.py
+sparam-conformance  corpus verification (stdlib only)
+
+  cases              11
+  physically valid   7
+  physically invalid 4
+  laws               5
+  sha256 matched     11/11
+
+  self-consistent: every file matches its recorded digest, and every
+  label agrees with itself
+
+  scope: internal consistency only. The digests ship in this download,
+  so this detects corruption, not substitution. It does not re-derive
+  the physics -- run score.py against a checker for that.
+```
+
+It verifies that every file matches the SHA-256 recorded in the index, that the
+index and manifest agree, and that no label contradicts itself. It exits 1 if
+any of that fails, so it is usable as a CI step. It deliberately does **not**
+re-derive the physics: a corpus that graded its own ground truth would be
+arguing with itself. That is what `score.py` is for. Nor is it an authenticity
+check — the digests ship inside the same download, so it catches corruption and
+inconsistent editing, not deliberate substitution.
+
 The corpus is committed, so scoring needs nothing installed beyond your own
 checker. `generate.py` is only needed if you want to prove the files came from
 the generator — it rebuilds them byte-identically, and CI checks that it does.
@@ -154,19 +185,19 @@ asymmetry for reciprocity, row power for energy. A corpus whose ground truth is
 wrong is worse than no corpus, because every checker scored against it inherits
 the error.
 
-That check earned its place: it caught **three wrong labels** during development.
-The "passive" resonator had σ_max = 1.2441 and was not passive at all; a case
-meant to isolate energy was also non-reciprocal; and `negative_resistance`
-unavoidably breaks energy conservation too, which the original label denied.
-
-That 1.2441 is the one figure on this page you cannot reproduce from the
-committed corpus — it belonged to a superseded generator revision, and the
-resonator that ships today has σ_max ≤ 1, which the test suite asserts. It is
-recorded because "we found bugs in our own ground truth" is worth more with a
-number attached than without one.
-
 Generation is deterministic, files are SHA-256 pinned in the manifest, and both
-properties are tested.
+properties are tested. `verify.py` re-checks those digests on your machine, so a
+corrupted or half-edited copy is caught before you score anything against it.
+
+The re-derivation is not ceremonial — it rejects labels. Three were corrected
+before release: a resonator labelled passive measured σ_max = 1.2441 and was
+not; a case meant to isolate an energy violation was also non-reciprocal; and
+`negative_resistance` unavoidably breaks energy conservation, which its first
+label denied. Every one was caught by the independent check rather than by a
+user, which is the entire argument for running it. (That 1.2441 belonged to a
+superseded generator revision; the resonator that ships has σ_max ≤ 1, asserted
+in the suite, so it is the one figure here you cannot reproduce from the
+committed files.)
 
 ## Scope, honestly