fix(ci): ship plugin content whenever it changes, and only when it changes

[TMHSDigital]

Closes #350.

Problem

Plugin updates ran backwards:

• Content fixes stuck: a docs: fix to a skill never cut a release, so it never reached plugin-dist. Examples: bce71a94, c9bb9988.
• Empty updates: every gallery or site fix: release rebuilt plugin-dist with identical content and a new version, prompting every user to update. v0.143.2 was site-only, and this session cut about a dozen more.

Change

• .github/scripts/plugin-content-changed.sh BASE: exits 0 when skills/ claude/ rules/ snippets/ templates/ differ between BASE and HEAD. release.yml cuts a patch release when this says yes and no commit subject is feat/fix.
• build_plugin_dist.py --fingerprint DIR: sha256 over every file, with manifest "version" values blanked and CRLF normalized.
• release.yml publish step: builds, fingerprints, fetches the published plugin-dist, and skips the force-push when the fingerprints match. The plugin version now moves only with its content.
• Docs and tests: AGENTS.md documents both rules. New tests/test_plugin_content_changed.py (temp git repo, 4 cases), wired into Validate.

Evidence (local)

• Fingerprint:
  • two builds of the same tree are equal;
  • changing only "version" keeps them equal;
  • appending one line to skills/operators/SKILL.md makes them differ.
• Against the live branch: today's build fingerprints identically to the published plugin-dist (plugin build v0.143.13). The raw diff -r shows differences, but diff -r --strip-trailing-cr leaves 0 differences outside "version" lines. So under this change, a site-only release would not republish.
• Detector: test_plugin_content_changed.py 4/4 OK. README and docs/gallery changes are not content; skill and snippet edits are.
• test_bump_kind.py and test_build_plugin_dist.py OK. Both workflow files parse.

Not exercisable locally: a real release run. The two new steps are plain shell around the tested script and CLI. The first release after merge will log plugin-dist fingerprint: built=… published=… as the observable check.

🤖 Generated with Claude Code

[TMHSDigital]

CI evidence before merge: all checks pass ( 13 pass ). Blender Smoke on PR head: Blender 5.2.2 LTS / Blender 4.5.14 LTS. Summaries (happy path; falsifiers) per leg: 141 passed, 3 skipped, 0 failed;136 passed, 7 skipped, 0 failed;143 passed, 1 skipped, 0 failed;143 passed, 0 skipped, 0 failed;