Revised 0.9 workflow comparison

The before program is the exact qualified f456958c example, run with its preserved ABI-53 wheel. The current program runs with the ABI-57 wheel. Both complete all five tasks in fresh fixtures. The original 0.8 comparison remains available.

Task	Nonblank lines, before → current	Caller work removed or retained
Import	8 → 9	Returns both identities; the additional assignment retains the asset ID for later tasks without another query
Adopt	9 → 8	One path lookup replaces fallible path-to-URI conversion and a locator wrapper; ambiguity and continuation checks remain
Relink	21 → 17	Verified selection carries evidence and the original base; removes first-member assumptions and copying locator fields
Output	27 → 15	One output/activity operation replaces separate staging and edge wrappers; uses the retained input ID, actual tool and parameter
Revisions	5 → 5	The explicit-view bounded scans remain useful and unchanged

Counts include comments and assertions. They are supporting evidence rather than an acceptance threshold. The before output also incorrectly recorded a provenance input as an authored dependency. Removing that assertion corrects meaning; it is not an API convenience saving. The current harness checks that no dependency observation was invented. A separate scene fixture demonstrates actual authored references.

Relinking now rejects partial verification, missing required members and incomplete search. Selection survives disposal of the native result and view; confirmation stages all locations against its original base. An existing location is a guarded no-op. Commit remains explicit, conflicts remain visible, and database guards do not guarantee unchanged external files.

No convenience operation adds a scan, subprocess or automatic retry. Paths convert once; import/output preparation hashes before taking the write edit; resolution releases its temporary production read before filesystem work. Explicit retained views still have caller-controlled lifetimes. Python metadata reads use scalars and immutable tuples, retaining tags for ambiguous meanings and ordered repeated fields. Jobs and exchange live in discoverable modules; their distinct ownership and deferred-intent capabilities are preserved.

All five surfaces have installed runnable recipes linked from the migration guide. Native contracts and Python regressions cover malformed inputs, incomplete/ambiguous selection, copied evidence lifetime, terminal conflicts, no-op receipts, exact metadata and effect capture/replay. Symbol counts and shorter snippets do not establish those semantics.