Make compare-workspace observationally pure with respect to the repository it
analyses. Its requested output directory must control both pipeline artifacts
and extractor caches, including when that directory is outside the repository.
Generated evidence must not appear in the captured Git status or in either
intent graph. A generated graph that exceeds the generic JSON-document ceiling
but remains within the comparison-specific resource budget must be readable by
the comparator without weakening limits for unrelated JSON consumers.
outputDir receives current-workspace artifacts and
caches without creating .intent in the analysed repository.compare-workspace accepts the observed 142,557,246-byte Platform
graph under an explicit 256 MiB ceiling while still rejecting graphs
above that bounded comparison limit.The user’s request to implement the missing ticket-to-code verification loop, followed by explicit continuation, authorizes this bounded defect repair. This conversation note is not merge authorization; protected exact-head Validator evidence remains required.
On 2026-08-25 the Subactor coding-agent independently reproduced the live
failure on PLF-8091; the Founder asked the supervisor to continue and repair
autonomization defects found through observation. Exact stderr proved that both
pipelines completed but the comparator rejected a 142,557,246-byte generated
graph at the generic 128 MiB read ceiling.
npm run verify passes 421 tests with one existing JDK-only skip; type,
module, environment, workflow, generated-analysis, structured-response and
schema gates pass.make docker-smoke, governance and diff checks pass locally. Protected
exact-head review remains pending.compare-workspace with a
142,557,246-byte generated graph and zero stderr. It then surfaced the real
semantic delta (blocking +2) instead of a file-size stack trace.Protected PRs #98 and #100 merged the isolated-output and bounded large-graph
repairs; both are ancestors of main@a93944f51d47fbf8fe1f3aaea03f17f4fb472d80.