feat(yield): convert-skill — the converter is itself a Yield skill

[bigboateng]

Closes the loop the whole lab exists for: existing prose skills get a paved road into Yield, in the language the operator picks (Go / TypeScript / Python / Rust) — and the converter is itself a Yield skill, running under the same supervisor it generates programs for.

Locus-decided design (yield/docs/locus-converter.md)

• supervisory-rw: controllable — shipping an unverified conversion is preventable; nonblocking — every conversion ends SHIPPED_VERIFIED or honestly REPORTED_BLOCKED.
• Decided rival-design comparison (safety-reachability): the executed-verification design satisfies forbidden-unreachable; completing on the model's transcription is rejected with the verbatim trace extract_flow → pick_language → generate_program → complete_untested → SHIPPED_UNVERIFIED.

The spine: a conversion that was never executed is never "done." The program owns pipeline order, the closed language menu, the ≤2 repair bound, and the evidence gate — a nested yskill test of the generated skill, executed by the supervisor as observed fact. The model owns flow extraction and code writing.

Contents

• examples/convert-skill (Go): ask for the source prose skill → read it (run_command, observed) → agent_task extract the implicit flow (schema-bound) → ask_user pick the target language → agent_task write the files → nested yskill test of the result, with a bounded repair loop → evidence-bound completion or honest Blocked.
• Scripted fixture exercises the full machinery including the nested test run (${YSKILL:-yskill}; CI pins YSKILL=/tmp/yskill).
• Locus models + rival-design derivation under docs/locus/; design doc docs/locus-converter.md; README pointer.

Verification

go vet clean; YSKILL=/tmp/yskill yskill test examples/convert-skill reaches completed locally (nested test of the destination skill included); CI runs it in the yield-lab validate job.

Also in this cycle (context, already merged/repaired)

The public-repo sync had wedged silently: the manual seed recorded a branch commit that never landed on main, so the stale-guard ignored every sync while reporting success. Repaired by re-projecting from main @ 7e6447d0 — this PR's merge is the first true end-to-end test of the sync loop.