deps: refresh measure-mirror pin → v0.27.0 + ㉗ declared-chance contract (v0.2.7)

CHANGELOG.md

@@ -5,6 +5,26 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
 
 ---
 
+## [0.2.7] — 2026-07-21
+
+### Changed
+- **measure-mirror pin `13077df` → `690c27e` (v0.26.0 → v0.27.0)**. Picks up, in order:
+  - **v0.27.0 — full 64-hex SHA-256 seals** (security, SPEC v1.1): 16-hex truncated
+    seals allowed a birthday-search (~2³² hashes) to forge two entries sharing one seal
+    and swap them post-seal. Legacy 16-hex seals keep verifying (prefix match) — no
+    ledger migration. The stack's L1 chain check inherits the wider digest.
+  - **v0.26.1 — ㉗ `prereg_lint` false-positive fix**: an audit of the probe against 64
+    real ledgers found ⑫c was reading `baseline` as the chance floor → **44 spurious
+    FAILs = 44 wrong compute-gate BLOCKs**. It now uses a declared `chance` only, so a
+    below-chance gate BLOCK requires `chance=` on the preregistration.
+- Docs: "bar at/below chance" → "bar at/below **declared** chance" (README, connect-time
+  DISCIPLINE, `mm_prereg_lint` reminder).
+
+### Fixed
+- **`test_compute_gate_blocks_on_lint_fail_below_chance_bar`** now declares `chance=`
+  (the corrected contract) + new `test_compute_gate_does_not_block_on_baseline_alone`
+  pins the false-positive guard: `pass < baseline` with no declared `chance` must GO.
+
 ## [0.2.6] — 2026-07-21
 
 ### Added

README.md

@@ -43,7 +43,7 @@ Windsurf, any MCP client.
 | `mm_audit` | 🪞 | audit a result vs its sealed registration |
 | `mm_power_check` | 🪞 | false-negative guard — is n big enough? (design-time) |
 | `mm_falsifiability_check` | 🪞 | Popper gate — kill-condition registered & not tripped? |
-| `mm_prereg_lint` | 🪞 | seal *quality* lint (㉗) — leaked kill-condition, bar at/below chance, unstructured kill, low n, undeclared pre-seal checks; the compute gate BLOCKs on a lint FAIL |
+| `mm_prereg_lint` | 🪞 | seal *quality* lint (㉗) — leaked kill-condition, bar at/below declared chance, unstructured kill, low n, undeclared pre-seal checks; the compute gate BLOCKs on a lint FAIL |
 | `mm_leakage_check` | 🪞 | train∩test contamination |
 | `mm_multiseed_check` | 🪞 | unstable signal / lucky seed |
 | `mm_retract` | 🪞 | chain-linked retraction (cannot be silently deleted) |

mirror_stack_mcp/__init__.py

@@ -1,2 +1,2 @@
 """🪞🔎🪪 Mirror Stack unified MCP server."""
-__version__ = "0.2.6"
+__version__ = "0.2.7"

mirror_stack_mcp/server.py

@@ -43,7 +43,7 @@
      neutral-control / manipulation-check / positive-control).
   2. Power: is n big enough to detect the effect? mm_power_check (design-time).
   3. Lint the seal — mm_prereg_lint (auto in the mm_preregister response). A FAIL
-     (kill-condition leaked into the metric field, bar at/below chance) means the
+     (kill-condition leaked into the metric field, bar at/below declared chance) means the
      automated checks can't fire: fix and re-seal under a NEW claim_id.
 
 BEFORE reporting (verify before speaking):
@@ -70,7 +70,7 @@
 NORMATIVE FORMAT: the ledger format is defined by MIRROR-SPEC v1.0 (measure-mirror
 docs/SPEC.md, ratified & frozen 2026-07-02) — these tools are a reference implementation.
 KNOWN ILLUSIONS: before trusting a surprising result (good OR bad), scan the catalog of
-measurement illusions (measure-mirror catalog/, 45 real sealed cases: gaming, self-catch,
+measurement illusions (measure-mirror catalog/, 46 real sealed cases: gaming, self-catch,
 false-negative guards, contamination). When you retract, consider adding a specimen.
 """
 
@@ -138,7 +138,7 @@ def _compact(findings):
     "mm_falsifiability_check": "🪞 If the kill-condition tripped (FAIL), the claim is falsified by "
         'its OWN criterion → mm_retract it. If it did not trip, that\'s "not refuted", not "proven".',
     "mm_prereg_lint": "🪞 Lint judges the seal's QUALITY, not just its presence. A FAIL here "
-        "(kill-condition leaked into `metric`, or a bar at/below chance) means the automated "
+        "(kill-condition leaked into `metric`, or a bar at/below declared chance) means the automated "
         "checks can't fire — fix and re-seal under a NEW claim_id (first-write-wins). Run this "
         "right before spending compute; it's the cheap machine-check that saves a KILL.",
     "mm_preflight": "🪞 This is a primitive — the MCP only judges GO/BLOCK. YOUR launcher / "

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mirror-stack-mcp"
-version = "0.2.6"
+version = "0.2.7"
 description = "Unified MCP server for the Mirror Stack — claims, actions, provenance + verify-all in one server"
 readme = "README.md"
 requires-python = ">=3.10"
@@ -16,7 +16,7 @@ dependencies = [
     # tracks each mirror's default branch, so an install made before a mirror's release
     # freezes on whatever was HEAD then and never refreshes — that is how a local install
     # ran stale measure-mirror 0.14.3 long after 0.15.1 landed. Bump these on purpose.
-    "measure-mirror @ git+https://github.com/bhyi4/measure-mirror@13077df8c2616e0559742b2d97c37e2459d7906c",   # v0.26.0 — ㉗ prereg_lint pre-compute seal-quality probe + pre_seal_checks (#28; required by mm_prereg_lint/gate lint wiring); v0.25.0 catalog v1.8 + anchor probes ㉔㉕ A2; v0.24.0 grounding ㉑㉒㉓ + A1
+    "measure-mirror @ git+https://github.com/bhyi4/measure-mirror@690c27e9645878576f7284bdcc9b7cafdf7c0f53",   # v0.27.0 — full 64-hex SHA-256 seals (security, SPEC v1.1); v0.26.1 ㉗ prereg_lint false-positive fix (declared-chance floor); v0.26.0 ㉗ prereg_lint probe + pre_seal_checks (required by mm_prereg_lint/gate wiring); v0.24~0.25 grounding/anchor probes ㉑–㉕
     "action-mirror @ git+https://github.com/bhyi4/action-mirror@fd46e90c3e3f0706792c261316f4699d7a6799d8",     # v0.1.0 (ledgers conform to MIRROR-SPEC v1.0)
     "provenance-mirror @ git+https://github.com/bhyi4/provenance-mirror@8edbbfd94772d3718de93271c1f80297c0476043",  # v0.1.0 (ledgers conform to MIRROR-SPEC v1.0)
 ]

tests/test_server.py

@@ -189,15 +189,27 @@ def test_preregister_response_carries_auto_lint(tmp_path):
 
 def test_compute_gate_blocks_on_lint_fail_below_chance_bar(tmp_path):
     led = tmp_path / "mm.jsonl"
-    # well-formed kill fields, but the pass bar sits below chance → lint FAIL
-    s.mm_preregister(str(led), "c1", metric="acc", baseline=0.5, pass_threshold=0.45,
+    # well-formed kill fields, but the pass bar sits at/below the DECLARED chance → FAIL.
+    # (chance must be declared — `baseline` alone is a comparison arm, not the floor.)
+    s.mm_preregister(str(led), "c1", metric="acc", chance=0.5, pass_threshold=0.45,
                      kill_threshold={"metric": "acc", "threshold": 0.4,
                                      "direction": "below"})
     r = s.mm_preflight(str(led), "c1", gate="compute")
     assert r["decision"] == "BLOCK"
     assert "prereg-lint FAIL" in " ".join(r["reasons"])
 
 
+def test_compute_gate_does_not_block_on_baseline_alone(tmp_path):
+    led = tmp_path / "mm.jsonl"
+    # pass=0.45 < baseline=0.5 but NO declared chance → must GO (A3 false-positive guard).
+    s.mm_preregister(str(led), "c1", metric="acc", baseline=0.5, pass_threshold=0.45,
+                     min_n=200,
+                     kill_threshold={"metric": "acc", "threshold": 0.4,
+                                     "direction": "below"})
+    r = s.mm_preflight(str(led), "c1", gate="compute")
+    assert r["decision"] == "GO"
+
+
 def test_compute_gate_go_passes_lint_check(tmp_path):
     led = tmp_path / "mm.jsonl"
     s.mm_preregister(str(led), "c1", metric="acc", min_n=240,