Release notes come from a file in the repo, not the tag

Deriving the body from the tag annotation does not survive CI: the
release action does not fall back to it (v1.2.0 published empty), and
checkout leaves the tag lightweight, so %(contents) yields the commit
message instead - which is what v1.2.1 first published. Keep the notes
at docs/release-notes/v<version>.md, pass that as body_path, and fail
the run when it is missing.

docs/RELEASING.md

@@ -4,7 +4,7 @@ The PCM addon zip is built by CI (`.gitea/workflows/build-pcm.yml`).
 Every push to `main` builds it as a downloadable artifact; pushing a
 `v<version>` tag additionally creates a Gitea release with the zip
 attached. The release job checks that the tag matches `metadata.json`
-and fails on a mismatch.
+and that the release notes exist, and fails on either mismatch.
 
 ## Steps
 
@@ -23,17 +23,25 @@ and fails on a mismatch.
    ]
    ```
 
-2. **Commit, tag, push** (tag = `v` + the manifest version):
+2. **Write the release notes** at `docs/release-notes/v<version>.md`.
+   This file becomes the release description verbatim; the job fails if
+   it is missing or empty (the release action publishes empty notes
+   rather than falling back to the tag message, which is how v1.2.0
+   shipped with a blank description). Say what changed for a user of
+   the previous version — in particular, whether results move for an
+   unchanged board.
+
+3. **Commit, tag, push** (tag = `v` + the manifest version):
 
    ```powershell
-   git add metadata.json
+   git add metadata.json docs/release-notes/v1.0.2.md
    git commit -m "Release 1.0.2"
    git tag v1.0.2
    git push
    git push origin v1.0.2
    ```
 
-3. **Verify**: the Actions run for the tag builds
+4. **Verify**: the Actions run for the tag builds
    `th.co.b4l.fill-resistance_<version>.zip` and publishes it at
    <https://git.b4l.co.th/B4L/kicad-zone-resistance/releases>, together
    with `metadata-registry.json`. The zip installs directly via

docs/release-notes/v1.2.1.md

@@ -0,0 +1,36 @@
+Bug-fix release. Results are unchanged from 1.2.0 for a board that
+solves cleanly; the fixes are in the in-KiCad overlay push, pad copper
+selection and error reporting.
+
+Note for anyone coming from 1.1.0 or earlier: 1.2.0 changed the physics
+model (exact SMD and THT pad copper, populated THT holes conducting as
+their solder plug and lead, slotted holes as true stadiums) and fixed an
+adaptive barrel-refinement bug that could make via-field results read up
+to ~13% low. Numbers for an unchanged board differ from 1.1.0 - re-run
+any board you track across versions.
+
+Fixed:
+- Overlay push: a locked reference image silently survived removal and a
+  new one was stacked on top of it. KiCad reports the failure per item
+  while the overall request still reads OK; it is now checked, and the
+  layer is reported and skipped instead.
+- Overlay push: a run covering fewer layers than the previous one left
+  the earlier solve's heatmap on the unused slots, where it read as
+  current. Those slots are now cleared.
+- Overlay push: the whole push is one commit, so a single undo reverts
+  it rather than just the last layer.
+- Through-hole pad copper was always read from F.Cu even when the joint
+  protrudes on B.Cu, mis-sizing the modelled solder coat for pads sized
+  differently per copper layer. The solder side is now probed first.
+- A failure before the output directory existed - a broken plugin
+  Python environment, typically - reported nothing at all on screen.
+  The error figure now falls back to the temp directory.
+- Pads sitting on no single copper layer are noted rather than silently
+  skipped, and the frequency field keeps its specific rejection reason
+  ("1,500" is a thousands separator, "-5" is negative) as the other
+  numeric fields already did.
+
+The KiCad overlay push remains experimental and opt-in (off by default).
+It writes reference images to User.9-User.12 and replaces what is on
+those layers.
+