@dzhechkov/p-replicator 1.5.17 → 1.6.0

package/templates/.claude/skills/goap-research-ed25519/scripts/test_signature_v3.py

@@ -0,0 +1,554 @@
+#!/usr/bin/env python3
+"""
+ADR-003 / ADR-005 Confirmation tests — the v3 signed message (T-1…T-8, T-17, T-19).
+
+    cd .../goap-research-ed25519/scripts && python3 -m unittest discover -s . -p 'test_*.py' -v
+
+The load-bearing test here is `test_trust_class_reclaim_breaks_v3_signature_discrimination`:
+it models the LIVE attack — sign ISSUER_SIGNED while the pin is active, revoke the
+pin at verify time, reclaim to SELF_ATTESTED, verify — because a pin that was never
+active would fail for the wrong reason and prove nothing.
+"""
+
+import copy
+import json
+import os
+import re
+import sys
+
+# Import NOTHING local before this line (stray __pycache__ reads as canonical drift).
+sys.dont_write_bytecode = True
+
+import unittest
+
+import ed25519_verifier as ev
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+
+
+def _pop(description="men with obesity (BMI >= 30) enrolled in a weight-loss trial"):
+    return {
+        "description": description,
+        "criteria": {
+            "bmi_min": {"op": ">=", "value": 30, "kind": "baseline",
+                        "verbatim": "men with obesity (BMI >= 30 kg/m2)",
+                        "locator": "[Methods, Participants]"},
+        },
+    }
+
+
+def _verifier():
+    v = ev.Ed25519Verifier()
+    v.generate_keypair()
+    return v
+
+
+def _fixture(name):
+    with open(os.path.join(HERE, name), "r", encoding="utf-8") as handle:
+        return json.load(handle)
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class MandatoryStudyPopulationTests(unittest.TestCase):
+    """T-1 (AM-1 / FR-1) — five factories x four malformed inputs = 20 refusals.
+
+    ASSERTS THE `TypeError`, never "the suite is green" (AM-13): giving the argument
+    a default keeps the whole suite green and makes only this test fail — which is
+    exactly why it is written this way.
+    """
+
+    def setUp(self):
+        self.v = _verifier()
+        import evidence_fetch as ef
+        self.record = ef.FetchRecord(
+            url="https://pubmed.ncbi.nlm.nih.gov/123", final_url="https://pubmed.ncbi.nlm.nih.gov/123",
+            status=200, sha256_body="a" * 64, bytes_len=10, fetched_at="2026-08-03T10:00:00Z",
+            witness=ef._FETCH_WITNESS)
+
+    def _factories(self):
+        return {
+            "create_signed_fact": lambda **kw: self.v.create_signed_fact(
+                claim="c", source_url="https://x.test", source_content="b", issuer="researcher", **kw),
+            "create_issuer_signed_fact": lambda **kw: self.v.create_issuer_signed_fact(
+                claim="c", source_url="https://x.test", source_content="b", issuer="researcher", **kw),
+            "create_fetched_fact": lambda **kw: self.v.create_fetched_fact(
+                claim="c", fetch_record=self.record, issuer="researcher", **kw),
+            "create_listing_fact": lambda **kw: self.v.create_listing_fact(
+                claim="c", source_url="https://x.test", reason="offline run", **kw),
+            "create_asserted_fact": lambda **kw: self.v.create_asserted_fact(claim="c", **kw),
+        }
+
+    def test_create_fact_without_study_population_raises(self):
+        cases = 0
+        for name, factory in self._factories().items():
+            with self.subTest(factory=name, case="omitted"):
+                with self.assertRaises(TypeError):
+                    factory()
+                cases += 1
+            for case, value in (("None", None), ("empty string", ""),
+                                ("present but meaningless", {"description": "   ", "criteria": {}})):
+                with self.subTest(factory=name, case=case):
+                    with self.assertRaises(ValueError):
+                        factory(study_population=value)
+                    cases += 1
+        self.assertEqual(cases, 20, "5 factories x 4 malformed inputs")
+
+    def test_study_population_blank_or_empty_raises(self):
+        """T-2 — the API-shape refusal, exercised through the fact-construction path."""
+        for bad in ({"description": "adults", "criteria": {}},
+                    {"description": "", "criteria": {"bmi": 1}},
+                    {"criteria": {"bmi": 1}},
+                    "adults"):
+            with self.subTest(value=bad):
+                with self.assertRaises(ValueError):
+                    self.v.create_asserted_fact(claim="c", study_population=bad)
+
+    def test_unstated_requires_reason(self):
+        with self.assertRaises(ValueError):
+            self.v.create_asserted_fact(claim="c", study_population={
+                "description": "population not stated", "criteria": {}, "unstated_reason": "   "})
+        fact = self.v.create_asserted_fact(claim="c", study_population={
+            "description": "population not stated", "criteria": {},
+            "unstated_reason": "the abstract never describes who was enrolled"})
+        self.assertTrue(self.v.verify_fact(fact).verified)
+
+    def test_every_new_fact_is_v3(self):
+        for name, factory in self._factories().items():
+            with self.subTest(factory=name):
+                fact = factory(study_population=_pop())
+                self.assertEqual(ev.fact_schema_version(fact), 3)
+                self.assertIsNotNone(fact.study_population)
+                if fact.trust_class == ev.TRUST_CLASS_ISSUER_SIGNED:
+                    # An issuer-signed fact verifies against the PINNED key, never the
+                    # embedded one — that is the whole point of the class, so the pin
+                    # must exist for this check to mean anything.
+                    self.v.registry.add("researcher", self.v.get_public_key_b64())
+                self.assertTrue(self.v.verify_fact(fact).verified)
+
+    def test_from_dict_stays_permissive(self):
+        """FR-1's other half: mandatory on CREATION, never on LOADING. A loader that
+        refuses a legacy record cannot report on it."""
+        legacy = _fixture("fixture_legacy_v1_fact.json")["fact"]
+        fact = ev.SignedFact.from_dict(legacy)
+        self.assertIsNone(fact.study_population)
+        self.assertEqual(ev.fact_schema_version(fact), 1)
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class V3SignatureCoverageTests(unittest.TestCase):
+    """T-3 / T-5 — the four new keys are UNDER the signature, both tamper directions."""
+
+    def setUp(self):
+        self.v = _verifier()
+
+    def _fact(self, **kw):
+        return self.v.create_listing_fact(claim="c", source_url="https://x.test",
+                                          reason="offline run", study_population=_pop(), **kw)
+
+    def test_tamper_study_population_breaks_v3_signature(self):
+        """T-3 (AM-2 / FR-13) — strip / add / swap, the same three-test shape the
+        evidence axis already uses."""
+        stripped = self._fact()
+        self.assertTrue(self.v.verify_fact(stripped).verified)
+        stripped.study_population = None
+        self.assertFalse(self.v.verify_fact(stripped).verified, "stripping it must break the signature")
+
+        swapped = self._fact()
+        swapped.study_population = _pop("a completely different cohort")
+        self.assertFalse(self.v.verify_fact(swapped).verified, "swapping it must break the signature")
+
+        added = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        self.assertTrue(self.v.verify_fact(added).verified)
+        added.study_population = _pop()
+        self.assertFalse(self.v.verify_fact(added).verified, "adding it to a v2 fact must break it")
+
+    def test_dispatch_tamper_table(self):
+        """ADR-003 §2's dispatch rows — MEASURED, and one row came back different from
+        what the ADR predicted.
+
+        What refuses a tampered fact is the KEY-SET/VALUE difference between the two
+        reconstructed texts, never the `schema` marker or the `schema_version` field
+        (both are self-description). This module has already been taught that lesson
+        once: ADR-002's first draft credited the v2 marker with closing the downgrade
+        attack and its own discrimination run refuted it.
+        """
+        rows = {}
+        edited = self._fact()
+        edited.trust_class = ev.TRUST_CLASS_ISSUER_SIGNED
+        rows["edit trust_class"] = edited
+
+        downgraded = self._fact()
+        downgraded.schema_version = 1
+        rows["schema_version = 1"] = downgraded
+
+        both_stripped = self._fact()
+        both_stripped.schema_version = None
+        both_stripped.study_population = None
+        rows["strip schema_version AND study_population (the real downgrade)"] = both_stripped
+
+        upgraded = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        upgraded.schema_version = 3
+        rows["schema_version added to a v2 fact"] = upgraded
+
+        for name, fact in rows.items():
+            with self.subTest(row=name):
+                self.assertFalse(self.v.verify_fact(fact).verified, f"{name} must not verify")
+
+    def test_stripping_schema_version_alone_is_a_no_op_not_an_attack(self):
+        """MEASURED CORRECTION to ADR-003 §2's predicted tamper table (reported, not
+        smoothed over — the ADR listed this row as "must fail").
+
+        It does NOT fail, and it should not: dispatch is by PRESENCE (D-5), so with
+        `study_population` still there the verifier recovers version 3, rebuilds the
+        identical text, and the signature matches. Nothing about the fact changed —
+        `schema_version` is not in the signed message, it is a convenience mirror of a
+        property already derivable from the fields. The attack it was supposed to model
+        is covered by the row above: stripping the FIELD as well really does downgrade
+        the reconstruction, and really does fail.
+        """
+        fact = self._fact()
+        fact.schema_version = None
+        self.assertEqual(ev.fact_schema_version(fact), 3, "presence-dispatch recovers the version")
+        self.assertTrue(self.v.verify_fact(fact).verified,
+                        "a no-op edit must not be reported as tampering")
+
+    def test_tamper_metadata_breaks_v3_signature(self):
+        """T-5 / AC-13 — `evidence_note` is the MANDATORY reason for a LISTING_ONLY
+        degradation; rewriting it hollows out the audit trail the axis exists for."""
+        fact = self._fact()
+        self.assertIn("evidence_note", fact.metadata)
+        fact.metadata["evidence_note"] = "fetched and verified"
+        self.assertFalse(self.v.verify_fact(fact).verified)
+
+        stripped = self._fact()
+        stripped.metadata.pop("evidence_note")
+        self.assertFalse(self.v.verify_fact(stripped).verified)
+
+    def test_tamper_confidence_breaks_v3_signature(self):
+        """A DIFFERENT protection than the min() clamp: it protects a consumer that
+        reads `fact.confidence` directly without re-running verify_fact()."""
+        fact = self._fact()
+        fact.confidence = 0.99
+        self.assertFalse(self.v.verify_fact(fact).verified)
+
+    def test_confidence_is_signed_as_a_fixed_width_string(self):
+        """Float repr differs across runtimes; a signed text two runtimes serialize
+        differently is a signature that fails for the wrong reason."""
+        fact = self._fact()
+        message = json.loads(ev.canonical_fact_message_v3(fact))
+        self.assertEqual(message["confidence"], "0.5000")
+        self.assertIsInstance(message["confidence"], str)
+
+    def test_unserializable_payload_refuses_to_sign(self):
+        """D-18 — a silently dropped key is an UNSIGNED key, so signing fails closed."""
+        with self.assertRaises(ValueError):
+            self.v.create_asserted_fact(claim="c", metadata={"bad": float("nan")},
+                                        study_population=_pop())
+        with self.assertRaises(ValueError):
+            self.v.create_asserted_fact(claim="c", study_population={
+                "description": "d", "criteria": {"bmi": float("inf")}})
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class TrustClassReclaimDiscriminationTests(unittest.TestCase):
+    """T-4 (AM-4) — the MANDATORY discrimination proof, modelled as the live attack."""
+
+    def test_trust_class_reclaim_breaks_v3_signature_discrimination(self):
+        signer = ev.Ed25519Verifier(auto_generate_keypair=True)
+        verifier = ev.Ed25519Verifier(trusted_issuers={
+            "nature.com": {"pubkey_b64": signer.get_public_key_b64(), "status": "active"}})
+
+        fact = signer.create_issuer_signed_fact(
+            claim="testosterone therapy showed no excess cardiovascular harm",
+            source_url="https://nature.com/articles/example", source_content="body",
+            issuer="nature.com", study_population=_pop())
+        before = verifier.verify_fact(fact)
+        self.assertTrue(before.verified, "signed while the pin was ACTIVE")
+        self.assertEqual(before.trust_class, ev.TRUST_CLASS_ISSUER_SIGNED)
+
+        # …the key is revoked AFTER signing. This sequencing is the point: a pin that
+        # was never active would fail for the wrong reason and prove nothing.
+        verifier.registry.add("nature.com", signer.get_public_key_b64(), status="revoked")
+        self.assertFalse(verifier.verify_fact(fact).verified, "a revoked pin refuses the honest fact")
+
+        reclaimed = copy.deepcopy(fact)
+        reclaimed.trust_class = ev.TRUST_CLASS_SELF_ATTESTED  # the attack, in a text editor
+        result = verifier.verify_fact(reclaimed)
+        self.assertFalse(result.verified,
+                         f"the reclaim must not verify (got verified={result.verified}, "
+                         f"confidence={result.confidence})")
+        self.assertEqual(result.confidence, 0.0,
+                         "pre-fix this returned confidence=0.60 on a revoked-pin fact")
+
+        # …and the reverse direction, for completeness (FR-11).
+        promoted = signer.create_signed_fact(
+            claim="c", source_url="https://nature.com/articles/example", source_content="b",
+            issuer="nature.com", study_population=_pop())
+        self.assertTrue(ev.Ed25519Verifier().verify_fact(promoted).verified)
+        promoted.trust_class = ev.TRUST_CLASS_ISSUER_SIGNED
+        self.assertFalse(verifier.verify_fact(promoted).verified,
+                         "SELF_ATTESTED -> ISSUER_SIGNED must break it too")
+
+    def test_legacy_reclaim_belt_is_bounded_and_honest(self):
+        """M3.7 — the v1/v2 belt catches the REVOKED-BUT-PINNED case (issuer IS signed
+        in v1/v2, so the lookup cannot be redirected) and openly does NOT catch the
+        unpinned-issuer case, which stays launderable to SELF_ATTESTED @ 0.60."""
+        signer = ev.Ed25519Verifier(auto_generate_keypair=True)
+        legacy = ev.SignedFact(
+            claim="c", source_url="https://x.test", source_hash="0" * 64, issuer="nature.com",
+            issuer_pubkey=f"ed25519:{signer.get_public_key_b64()}", signature="",
+            timestamp="2026-08-03T12:16:18.191716Z", confidence=0.60,
+            trust_class=ev.TRUST_CLASS_SELF_ATTESTED)
+        legacy.signature, _ = signer.sign_content(ev.canonical_fact_message_v1(legacy))
+        self.assertEqual(ev.fact_schema_version(legacy), 1)
+
+        unpinned = ev.Ed25519Verifier()
+        residual = unpinned.verify_fact(legacy)
+        self.assertTrue(residual.verified, "the unclosable residual, named rather than hidden")
+        self.assertEqual(residual.confidence, 0.60)
+
+        pinned = ev.Ed25519Verifier(trusted_issuers={
+            "nature.com": {"pubkey_b64": signer.get_public_key_b64(), "status": "revoked"}})
+        caught = pinned.verify_fact(legacy)
+        self.assertFalse(caught.verified)
+        self.assertIn("revoked", caught.error)
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class BackwardCompatibilityTests(unittest.TestCase):
+    """T-6 / T-7 (AM-5) — COMMITTED fixtures, replayed. Comparing two CURRENT code
+    paths would mask exactly the regression these exist to catch."""
+
+    def test_v1_legacy_fact_still_verifies_after_v3(self):
+        fact = ev.SignedFact.from_dict(_fixture("fixture_legacy_v1_fact.json")["fact"])
+        self.assertEqual(ev.fact_schema_version(fact), 1)
+        self.assertTrue(ev.Ed25519Verifier().verify_fact(fact).verified)
+
+    def test_v2_legacy_fact_still_verifies_after_v3(self):
+        """The v2 fixture was generated from PRE-CHANGE code and committed (M0.2):
+        generating it after the first edit would have made this test self-fulfilling."""
+        fact = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        self.assertEqual(ev.fact_schema_version(fact), 2)
+        self.assertIsNotNone(fact.evidence_class)
+        self.assertIsNone(fact.study_population)
+        self.assertTrue(ev.Ed25519Verifier().verify_fact(fact).verified)
+
+    def test_v1_and_v2_messages_are_byte_frozen(self):
+        """D-4 / D-4a — the two older texts must not gain a key when v3 lands."""
+        v1 = ev.SignedFact.from_dict(_fixture("fixture_legacy_v1_fact.json")["fact"])
+        v2 = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        self.assertEqual(
+            sorted(json.loads(ev.canonical_fact_message_v1(v1))),
+            ["claim", "issuer", "research_context", "source_hash", "source_url", "timestamp"])
+        self.assertEqual(
+            sorted(json.loads(ev.canonical_fact_message_v2(v2))),
+            ["claim", "evidence_class", "fetch_date", "issuer", "research_context", "schema",
+             "source_date", "source_hash", "source_url", "timestamp"])
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class SignedFieldsTests(unittest.TestCase):
+    """T-8 (D-6) — the pre-v3 hole becomes a VALUE a caller can branch on."""
+
+    def test_verification_result_names_unsigned_fields(self):
+        v1 = ev.SignedFact.from_dict(_fixture("fixture_legacy_v1_fact.json")["fact"])
+        v2 = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        verifier = ev.Ed25519Verifier()
+
+        r1 = verifier.verify_fact(v1)
+        self.assertEqual(r1.schema_version, 1)
+        self.assertNotIn("trust_class", r1.signed_fields)
+
+        r2 = verifier.verify_fact(v2)
+        self.assertEqual(r2.schema_version, 2)
+        self.assertNotIn("trust_class", r2.signed_fields)
+        self.assertIn("evidence_class", r2.signed_fields)
+
+        signer = _verifier()
+        v3 = signer.create_asserted_fact(claim="c", study_population=_pop())
+        r3 = signer.verify_fact(v3)
+        self.assertEqual(r3.schema_version, 3)
+        for key in ("trust_class", "metadata", "confidence", "study_population"):
+            self.assertIn(key, r3.signed_fields)
+
+
+class TierCeilingTests(unittest.TestCase):
+    """T-19 (AM-15 / ADR-005 / D-20) — the ceiling survives the schema bump."""
+
+    @unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+    def test_am15_tier_ceiling_applies_to_schema_3_facts(self):
+        """Three assertions, one property.
+
+        Minting schema 3 must NOT switch the third ceiling off. The predicate used to
+        read `fact_schema_version(fact) != 2` — "applies to exactly v2" — so every fact
+        this slice creates would have scored 0.60 instead of the promised 0.40 from an
+        unknown domain, silently. The tier-A row proves the tier term is actually being
+        MEASURED rather than being a constant; the v1 row proves the lower edge (no
+        retroactive re-scoring) still holds.
+        """
+        v = _verifier()
+        unknown_domain = v.create_fetched_fact(
+            claim="c", issuer="researcher", study_population=_pop(),
+            fetch_record=_authentic_record("https://some-blog.example/p"))
+        self.assertEqual(ev.fact_schema_version(unknown_domain), 3)
+        self.assertEqual(v.verify_fact(unknown_domain).confidence, 0.40,
+                         "tier D must cap a schema-3 fact, exactly as it caps a schema-2 one")
+
+        tier_a = v.create_fetched_fact(
+            claim="c", issuer="researcher", study_population=_pop(),
+            fetch_record=_authentic_record("https://www.cochrane.org/x"))
+        self.assertEqual(v.verify_fact(tier_a).confidence, 0.60,
+                         "tier A leaves the trust ceiling binding — the tier term is measured, not constant")
+
+        legacy = ev.SignedFact(
+            claim="c", source_url="https://some-blog.example/p", source_hash="0" * 64,
+            issuer="researcher", issuer_pubkey=f"ed25519:{v.get_public_key_b64()}", signature="",
+            timestamp="2026-08-03T12:16:18.191716Z", confidence=0.60,
+            trust_class=ev.TRUST_CLASS_SELF_ATTESTED)
+        legacy.signature, _ = v.sign_content(ev.canonical_fact_message_v1(legacy))
+        self.assertEqual(ev.fact_schema_version(legacy), 1)
+        self.assertEqual(v.verify_fact(legacy).confidence, 0.60,
+                         "v1 stays exempt — old records keep the semantics they were created under")
+
+    def test_the_scope_is_a_lower_bound_not_an_equality(self):
+        """The shape of the fix, asserted directly: any FUTURE schema must inherit the
+        ceiling without another migration. An equality gate has to be re-edited by
+        every migration; a lower bound is edited once."""
+        self.assertEqual(ev.TIER_CEILING_MIN_SCHEMA, 2)
+        source = open(os.path.join(HERE, "ed25519_verifier.py"), encoding="utf-8").read()
+        self.assertNotIn("fact_schema_version(fact) != 2", source,
+                         "the equality predicate must not return")
+        self.assertIn("< TIER_CEILING_MIN_SCHEMA", source,
+                      "the scope must stay a `<` lower bound against the named constant")
+
+
+@unittest.skipIf(ev.CRYPTO_BACKEND is None, "No Ed25519 backend installed")
+class SchemaVersionRefusalTests(unittest.TestCase):
+    """QE G6 — a schema the verifier cannot identify is REFUSED, never crashed on.
+
+    MEASURED before the fix: loading the committed v2 fixture with
+    `schema_version='not-a-number'` and calling `verify_fact` raised an UNCAUGHT
+    `ValueError` out of `int(fact.schema_version)`, taking down a gate run with a
+    traceback instead of returning `verified=False`. And within the old `>= 3` band the
+    value was not covered by the signed text, so `3` could be moved to `99` and the
+    fact still verified.
+    """
+
+    def _v3_fact(self):
+        v = _verifier()
+        return v, v.create_listing_fact(claim="c", source_url="https://x.test",
+                                        reason="offline run", study_population=_pop())
+
+    def test_malformed_schema_version_refuses_instead_of_raising(self):
+        fact = ev.SignedFact.from_dict(_fixture("fixture_legacy_v2_fact.json")["fact"])
+        verifier = ev.Ed25519Verifier()
+        self.assertTrue(verifier.verify_fact(fact).verified, "the honest fixture still verifies")
+
+        fact.schema_version = "not-a-number"
+        result = verifier.verify_fact(fact)          # must NOT raise
+        self.assertFalse(result.verified)
+        self.assertIn("not an integer", result.error)
+        self.assertEqual(result.schema_version, ev.SCHEMA_VERSION_UNIDENTIFIED,
+                         "an unidentifiable record is not reported as legacy v1")
+
+    def test_every_malformed_spelling_is_refused(self):
+        for bad in ("not-a-number", "", "3.5", "  ", [], {}, True, 2.9, object()):
+            with self.subTest(value=bad):
+                self.assertIsNone(ev.coerce_schema_version(bad))
+        for good, expected in ((3, 3), ("3", 3), (" 2 ", 2), (3.0, 3)):
+            with self.subTest(value=good):
+                self.assertEqual(ev.coerce_schema_version(good), expected)
+
+    def test_moving_the_version_inside_the_old_band_now_breaks_verification(self):
+        """`3 → 99` used to verify: the band accepted anything at or above 3 as v3."""
+        v, fact = self._v3_fact()
+        self.assertTrue(v.verify_fact(fact).verified)
+        fact.schema_version = 99
+        result = v.verify_fact(fact)
+        self.assertFalse(result.verified, "an unknown schema must not be treated as v3")
+        self.assertIn("99", result.error)
+        self.assertEqual(ev.KNOWN_SCHEMA_VERSIONS, (1, 2, 3))
+
+    def test_an_unknown_schema_has_no_message_to_reconstruct(self):
+        v, fact = self._v3_fact()
+        fact.schema_version = 4
+        with self.assertRaises(ev.SchemaVersionError):
+            ev.canonical_fact_message(fact)
+        self.assertEqual(ev.signed_fields_for(fact), (),
+                         "and no field may be claimed as signed for it")
+
+    def test_the_gate_reports_a_malformed_record_instead_of_dying(self):
+        """The composition that motivated the finding: this ran inside a gate."""
+        import check_report_evidence as gate
+        fact = dict(_fixture("fixture_legacy_v2_fact.json")["fact"], schema_version="not-a-number")
+        findings = gate.verify_ledger_signatures([fact])
+        self.assertTrue(findings, "a record the verifier refuses must surface as a finding")
+        self.assertIn(findings[0].kind, ("TAMPERED_FACT", "UNVERIFIABLE_FACT"))
+
+
+def _authentic_record(url):
+    import evidence_fetch as ef
+    return ef.FetchRecord(url=url, final_url=url, status=200, sha256_body="a" * 64,
+                          bytes_len=10, fetched_at="2026-08-03T10:00:00Z", witness=ef._FETCH_WITNESS)
+
+
+class GrepGateTests(unittest.TestCase):
+    """T-17 — no module builds a canonical fact message outside the three functions."""
+
+    def test_no_inlined_canonical_message_dict(self):
+        """Step-0 recalled pattern #2: a new resolver does not deliver a migration
+        unless every call site is wired — so grep for old/duplicate literals."""
+        offenders = []
+        for name in ("population_match.py", "risk_statement.py", "check_report_evidence.py"):
+            text = open(os.path.join(HERE, name), encoding="utf-8").read()
+            for match in re.finditer(r'"claim"\s*:', text):
+                window = text[max(0, match.start() - 400):match.start() + 400]
+                if '"source_hash"' in window or '"issuer_pubkey"' in window:
+                    offenders.append(f"{name}:{text[:match.start()].count(chr(10)) + 1}")
+        self.assertEqual(offenders, [], "a canonical message is built in ed25519_verifier.py only")
+
+    def test_verifier_never_imports_the_semantics_module(self):
+        """05_architecture §1.1c — a crypto module whose correctness depends on
+        importing a semantics module repeats the fail-open shape source_tier_ceiling
+        was already bitten by."""
+        source = open(os.path.join(HERE, "ed25519_verifier.py"), encoding="utf-8").read()
+        self.assertNotIn("import population_match", source)
+        self.assertNotIn("from population_match", source)
+
+    def test_no_new_file_assumes_an_install_layout(self):
+        """M7.2 — keeps the eventual cross-copy propagation cheap.
+
+        The needles are assembled at runtime so that this file, which must scan
+        ITSELF, does not fail on its own assertion text.
+        """
+        needles = ("base" + "/skills", "templates" + "/.claude")
+        for name in ("population_match.py", "risk_statement.py", "test_signature_v3.py",
+                     "test_population_match.py", "test_risk_absolute.py",
+                     "fixtures_field_cases.json", "fixture_legacy_v2_fact.json"):
+            with open(os.path.join(HERE, name), encoding="utf-8") as handle:
+                text = handle.read()
+            for needle in needles:
+                with self.subTest(file=name, needle=needle):
+                    self.assertNotIn(needle, text)
+
+
+class SkillDocumentationTests(unittest.TestCase):
+    """T-16 (FR-15 / AC-15) — a WEAK test on purpose: it proves the sentence exists,
+    never that it is true. Documentation is layer 4 and this file says so."""
+
+    def test_skillmd_names_the_pre_v3_signature_hole(self):
+        path = os.path.join(os.path.dirname(HERE), "SKILL.md")
+        text = open(path, encoding="utf-8").read()
+        self.assertIn("can be rewritten without invalidating the signature", text)
+        self.assertIn("fact-v3", text)
+        self.assertIn("out of scope for this slice", text)
+
+    def test_skillmd_states_the_relative_risk_rule(self):
+        path = os.path.join(os.path.dirname(HERE), "SKILL.md")
+        text = open(path, encoding="utf-8").read()
+        self.assertIn("Relative risk never travels alone", text)
+        self.assertIn("BASELINE RISK NOT ESTABLISHED", text)
+
+
+if __name__ == "__main__":
+    unittest.main()

package/templates/.claude/skills/goap-research-ed25519/scripts/test_suite_completeness.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+"""
+T-20 / D-21 (AM-14, ADR-006) — every `test_*.py` on disk is actually COLLECTED.
+
+    cd .../goap-research-ed25519/scripts && python3 -m unittest discover -s . -p 'test_*.py' -v
+
+Why this file exists: before slice C, `test_ed25519_verifier.py` — the security suite
+of the module the slice edits — was named by NO run command anywhere. A report of
+"full suite green" would have been true of the command and false of the code. A test
+module that nothing runs is a file, not a check; this makes forgetting one fail BY
+NAME instead of by omission.
+"""
+
+import glob
+import os
+import sys
+
+sys.dont_write_bytecode = True
+
+import unittest
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+PATTERN = "test_*.py"
+
+
+def _on_disk():
+    return sorted(os.path.basename(p)[:-3] for p in glob.glob(os.path.join(HERE, PATTERN)))
+
+
+def _collected():
+    names = set()
+
+    def walk(suite):
+        for item in suite:
+            if isinstance(item, unittest.TestSuite):
+                walk(item)
+            elif isinstance(item, unittest.loader._FailedTest):
+                # A module discovery could not IMPORT is reported BY NAME, never
+                # silently skipped — a gate that cannot run has cleared nothing.
+                names.add(item.id().split(".")[-1])
+            else:
+                names.add(type(item).__module__)
+
+    walk(unittest.TestLoader().discover(start_dir=HERE, pattern=PATTERN, top_level_dir=HERE))
+    return sorted(names)
+
+
+class SuiteCompletenessTests(unittest.TestCase):
+    def test_discovery_collects_every_test_file_on_disk(self):
+        on_disk = _on_disk()
+        collected = _collected()
+        missing = [name for name in on_disk if name not in collected]
+        self.assertEqual(missing, [],
+                         f"these modules exist but discovery does not collect them: {missing}")
+        self.assertGreaterEqual(len(on_disk), 6, "the directory must still hold its test modules")
+
+    def test_the_security_suite_of_the_edited_module_is_in_the_run(self):
+        """The specific omission AM-14 caught, pinned by name so it cannot recur."""
+        collected = _collected()
+        for required in ("test_ed25519_verifier", "test_evidence_provenance", "test_goap_planner",
+                         "test_signature_v3", "test_population_match", "test_risk_absolute"):
+            self.assertIn(required, collected)
+
+    def test_no_enumerated_run_command_is_reintroduced(self):
+        """ADR-006 D5's Monitoring clause: a SUITE command must be the discovery one,
+        so a hand-kept module list can never drift from the directory.
+
+        Scope, stated precisely: TWO OR MORE module names after `-m unittest` is an
+        enumeration and fails. A single module (`python3 -m unittest test_x -v`) is a
+        run-just-this-file hint, not a claim about the suite, and is allowed. This
+        file is skipped — it is the checker, and its own prose describes the pattern.
+        """
+        offenders = []
+        for name in sorted(os.listdir(HERE)):
+            if not name.endswith((".py", ".md")) or name == os.path.basename(__file__):
+                continue
+            with open(os.path.join(HERE, name), encoding="utf-8", errors="ignore") as handle:
+                text = handle.read()
+            for line in text.splitlines():
+                if "-m unittest" not in line or "discover" in line:
+                    continue
+                modules = [token for token in line.split("-m unittest")[1].split()
+                           if token.startswith("test_")]
+                if len(modules) > 1:
+                    offenders.append(f"{name}: {line.strip()}")
+        self.assertEqual(offenders, [], f"enumerated run commands found: {offenders}")
+
+
+if __name__ == "__main__":
+    unittest.main()

package/templates/.claude/skills/requirements-validator/SKILL.md

@@ -64,6 +64,13 @@ Always flag these terms and suggest specific replacements:
 
 **Score <50 = BLOCKED from development.** Provide rewrite suggestions.
 
+**Blocking floor — the weakest link decides, never the average.** Independently of the score, a
+requirement is BLOCKED if `Testable = 0` (no acceptance criteria) or `Completeness = 0` (rubric "No
+AC"). A story can total 72/100 with neither, and 72 reads as "minor fixes" in the table above. A non-zero `Testable` or `Completeness` REQUIRES quoting the acceptance criteria it scores — no quote
+means 0, because the agent that scores is the agent the floor binds. The
+floor is closed at those two: `Measurable` and `Traceability` are deliberately excluded — see
+`references/scoring-system.md` → "Blocking floor" for the worked case and the reason.
+
 ## Output Format
 
 ### Requirements Analysis Report