The test-branch release: everything on release/v3.2.0 after a week of local use, plus what that week surfaced. Thanks to @notbucki, @Nikhiltogadiya, @dispather, @y4r4i and @abhineshkr — one merged pull request and five issues resolved, with @notbucki's multi-log test run feeding straight back into the branch.
Multi-log aggregate review (#85, @notbucki). One review across several observation logs, opt-in and unchanged for a single log: absolute workspace roots, per-workspace activation tiers, PENDING.md created where absent, anchor workspace distinguished from anchor directory.
Frontmatter that parses (#81). The template quotes every prose value, and the session-start scan reports a suspect-header count beside the file count. Measured on a real 292-file log: 27 files failed yaml.safe_load, 26 caught, zero false positives.
Slugs keep non-Latin titles (#80). Latin diacritics fold, word characters of any script survive (NFC-normalised), and the migration report warns when more than a fifth of a corpus collapses to untitled.
Review offer and logging discipline (#90, #71, #58). The review offer scales its content with the backlog rather than its frequency; How to Log checks for a restatement before every write; resolutions name the staged path and the reconciliation gate escalates a copy that reaches its second review un-installed; Session Start step 4 says plainly what it cannot detect, and the weekly review gains the reachable half — a regression check for an activation tier that was present and is gone.
Starter set for existing installs. The starter-principles offer now reaches installs that predate it, via a marker file and a version line, and the set gains two entries: the coverage-boundary principle for companion skills and the bundle-and-progressive-disclosure principle.
Release-branch flow. A branch merges only when every change on it has either natural exercise recorded or a synthetic check run in the review with a stated pass criterion; the session that cuts a branch writes that test plan.
Session start and snippets. The scan appends a trace line to checkpoints.log, so a session that loaded the skill but never ran the protocol is distinguishable from one that did both; the archival sweep never word-splits a workspace path containing a space; both snippet loops state their bash requirement; the checkpoint flush also fires on a permission-stream failure.
Skill authoring. Progressive disclosure from the first draft regardless of size, with a two-mode relocation check; families (shared tool, or core plus companions) distinguished from pattern groups (shared shape only), and pattern-group members never name each other.
Full detail in the commit history; each commit names its issue and reporter.