v0.17.3 — findings from exercising the delivery path for real

Everything here came out of running the backlog-delivery path for real rather than reading it: the marker contract against a live /product:generate-frontend scaffold, the Output Location interlock on a scratch repository, the Step 1 sub-agent delegation, and the full tracker write path — status transitions, progress comments, PR linkage, merge, close, roll-up — against a throwaway repository instead of live work items.

Fixed

/architect:merge-issue — parent completion now reaches the manifest. Step 4 said "update the node", singular, alongside pr.merged and a merge SHA — fields that only exist on an Issue. A Sub-Epic or Epic transitioned to status::done on the tracker therefore had nothing written back. Observed exactly that in a real 35-node backlog: all 27 Issue nodes carried impl, while the Epic and four completed Sub-Epics had no impl key at all — so a later deliver-backlog resume, which reads impl.status, would see a finished Sub-Epic as unstarted. Step 4 now updates every node the roll-up moved, with pr fields scoped to the Issue.

/architect:export-backlog, /architect:deliver-backlog — the manifest's labels array is a creation seed, not live state. It records what was attached at creation and is never advanced; status lives on the tracker and in impl.status. In the real backlog every node still read status::todo while the tracker held 7 done and 9 doing. Harmless today because nothing reads it, but a trap for resume logic — so export-backlog states what the field is, and deliver-backlog is explicit that state comes from impl.status and the tracker, never from labels, with the tracker winning on disagreement.

/architect:generate-docs — the inventory digest separates observed from inferred. Exercising the Step 1 delegation validated the design — a haiku Explore agent returned a digest covering all six fields, matching independently verified ground truth on every count and command, with no source pasted and nothing fabricated. It also exposed a gap: the digest asserted "Node.js 18+" as inventory when package.json declares no engines — an inference from Vite 5's requirement. The conclusion was right, but the orchestrator holds only the digest and cannot tell inference from observation, so it reaches the README as fact, defeating the skill's central discipline. Derived values must now be marked inferred: <value> (<basis>).

/product:generate-frontend — regeneration says so before it overwrites. The output location is correct (the tree is meant to be replaced, which is why adapt-change re-runs the skill), but nothing warned that a re-run destroys hand-edits under the output root. It now states plainly that it will overwrite and confirms, offers --out=<path> to regenerate alongside instead of over, and notes that a scaffold which has graduated into a hand-maintained frontend belongs outside generated/.

Added

skills/implement-backlog/output-location.test.sh — the Output Location interlock asserted as behaviour on a scratch repository: check-ignore rejects a source root under generated/ and names the matching rule, accepts a real services/ root inside the worktree, a docs commit lands on feature/<issue-id>-<slug> with the Issue reference and stages the intended file, and an output path git ignores stages nothing so the commit is refused rather than silently empty. 11 checks, exit 1 on failure.

The repo now carries two contract tests:

python3 skills/generate-docs/marker-mechanics.test.py     # 17 checks
bash    skills/implement-backlog/output-location.test.sh  # 11 checks

Changed

Both changelogs' link-reference blocks were stale — CHANGELOG.md stopped at 0.8.2, leaving 13 tagged versions unlinked plus a broken [0.7.0] link to a tag that was never cut, and CHANGELOG_ja.md had none at all. Both now carry one reference per existing tag. docs/codex-gap-analysis_ja.md said 80 skills; it has 87 (architect 50 / product 26 / scalardb 11).

Full changelog: CHANGELOG.md · CHANGELOG_ja.md