diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index b0d15af..c6ce62a 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -18,11 +18,16 @@ jobs:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-latest]
         # The full requires-python range on every OS, floor (3.10) to
-        # free-threaded 3.14 ("3.14t" is uv's key for the cp314t build; the
-        # whole dep tree ships cp314t wheels for all three OSes — verified
-        # against PyPI before widening the matrix).
+        # free-threaded 3.14 ("3.14t" is uv's key for the cp314t build).
         python: ["3.10", "3.11", "3.12", "3.13", "3.14", "3.14t"]
     runs-on: ${{ matrix.os }}
+    # windows + 3.14t is a DOCUMENTED upstream gap, not our code (2026-07-18):
+    # mcp requires pywin32 on win32, and pywin32==312 ships no cp314t wheel, so
+    # `uv sync` fails before our suite can even run. Linux/macOS 3.14t legs are
+    # strict and green. This cell is allowed to fail so the matrix reports
+    # reality without going red on someone else's wheel gap — and it starts
+    # passing automatically the day pywin32 publishes cp314t wheels.
+    continue-on-error: ${{ matrix.os == 'windows-latest' && matrix.python == '3.14t' }}
     steps:
       - uses: actions/checkout@v4
       - name: Install uv
diff --git a/README.md b/README.md
index 01d24a9..b8f70d0 100644
--- a/README.md
+++ b/README.md
@@ -123,7 +123,11 @@ Claude Code and OpenAI Codex setup (both stdio MCP): see [`docs/clients.md`](doc
 
 ## Status
 
-**v1.2.0 (pre-release, in review)** — the Development Observatory release:
+**v1.2.0 (pre-release, in review)** — the Development Observatory release
+(Python 3.10–3.14 incl. free-threaded 3.14t, CI-verified on Linux/macOS/Windows;
+one known upstream gap: *Windows* 3.14t cannot install yet because `mcp`'s
+`pywin32` dependency ships no cp314t wheel — that CI cell is a documented
+allowed-failure until pywin32 catches up):
 seven new backends extend the digest matrix from profilers to the whole
 edit→build→verify→measure loop. BuildDigest: `clang_time_trace` (compiler-phase
 traces, with `[total]` aggregate tagging), `cmake_profile` (configure-step
@@ -137,7 +141,6 @@ previous-green comparison proven on real two-run fixtures). RepoState:
 honest absence, never `0.0`). Zero new tools — every backend rides the same
 seven; 206 tests. Every fixture in the release was produced by the real tool it
 mimics — several briefs were corrected by what the artifacts actually contained.
-Python 3.10–3.14 including free-threaded 3.14t, Linux/macOS/Windows CI-verified.
 
 **v1.1.1** — contract-hardening hotfix (issues #2/#4/#5): registry format
 collisions are detected on the normalized key (a case-variant format can no