@@ -0,0 +1,72 @@
+<!-- SPDX-License-Identifier: MIT -->
+
+# PR #32 — CI correction and README refresh
+
+## Observed failure
+
+Both initial CI runs at `a6144f4` failed only the complete firmware host suite:
+[pull-request run](https://github.com/PyBLE-dev/PyBLE/actions/runs/34618742875)
+and [branch-push run](https://github.com/PyBLE-dev/PyBLE/actions/runs/34618673864).
+Each ran 1,947 tests with one error and 24 skips. The failing test was
+`test_visible_replacement_during_callback_is_rejected_and_preserved` in
+`test_v061_hil_preflight_hardening.py`: its replacement file had disappeared.
+The firmware shell phase and Android WebView checks passed. Firmware builds
+were skipped because their host-test prerequisite failed; they did not report
+compiler failures.
+
+The host evidence writer closed its original descriptor before invoking the
+validation callback. Linux could reuse that unlinked file's inode number for a
+replacement. Cleanup then mistook the replacement for its own output and
+removed it. The same lifetime gap existed in the shared v0.6.0-profile writer.
+
+## Correction and boundaries
+
+Both writers now retain the original descriptor until publication checks and
+failure cleanup finish. The inode remains reserved while ownership is checked.
+Existing replacement files, including replacements with identical bytes, are
+rejected without deletion. Descriptor cleanup still occurs on success/failure.
+
+Specification commit `8c7a1ff`, regression commit `0bde1e6`, and implementation
+commit `8771c8b` preserve the specification-first red/green sequence. New tests
+observe real descriptor lifetime without relying on filesystem inode-reuse
+timing. They failed before the correction on this macOS host as well.
+
+Only host qualification tools changed. Firmware runtime, native modules,
+board overlays and version pins still match published source `c8f549e` exactly.
+No release threshold was relaxed, failed evidence rewritten, board operated,
+firmware rebuilt or live website changed.
+
+## Verification before the update push
+
+- Three deterministic inode-lifetime test methods passed for both writers,
+  including success, failure cleanup and same-byte replacement cases.
+- The affected five-file writer/gate suite passed all 80 tests under Python
+  3.12, including the original failing regression.
+- All 42 publication tests and the Markdown link gate passed.
+- All six firmware roadmap/ledger tests passed after correcting the old
+  expectation that v0.6.0 must still be the current public release.
+- Website tests: 342 passed across 23 files.
+- Clean-room, whitespace and PR-range DCO checks passed.
+
+The full Python 3.12 host suite is also rerun independently. Its final outcome
+and the subsequent Ubuntu CI results belong to the PR checks; focused host
+tests above are not hardware qualification evidence.
+
+## README and issue #24
+
+The README now links the published v0.6.1 descriptor, exact source, owner
+confirmation, recovery material, tutorials, diagram and official examples.
+It documents the editable repository URL, branch selection, writable-folder
+guidance, multi-file deletion, App source version and offline privacy policy.
+Historical v0.6.0 citation metadata remains historical. Prepared store artifacts
+are not advertised as proof of a live release or later source changes.
+
+[Issue #24](https://github.com/PyBLE-dev/PyBLE/issues/24) describes the original
+automated release-admission checklist, including five hardening results, ten
+storage receipts and ten physical App/profile rows. The
+[owner-confirmed publication](firmware-v0.6.1-publication-2026-09-11.md) is recorded,
+but its retained descriptor still has five pending automated HIL rows. This PR
+does not certify all of #24's original criteria and must not auto-close it as
+completed. The maintainer may instead explicitly close that original workflow
+as superseded by the owner-confirmed publication decision, preserving the
+distinction. This follow-up neither closes the issue nor resumes hardware tests.