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

.github/workflows/yield-lab.yml

@@ -41,6 +41,7 @@ jobs:
           /tmp/yskill test examples/release-checklist
           /tmp/yskill test examples/env-doctor
           /tmp/yskill test examples/data-migration
+          YSKILL=/tmp/yskill /tmp/yskill test examples/convert-skill
 
   # Shift-left projection gate: materialize the public surface exactly as
   # operatorstack/yield would receive it and compile+test it, so a broken

labs/22-yield/public-readme/README.md

@@ -59,6 +59,12 @@ four languages and asserts identical observable protocol behavior.
 Non-Go skills declare their runner in `skill.json`:
 `{"run": ["node", "main.ts"]}`.
 
+Already have prose skills? `examples/convert-skill` is a converter —
+itself a Yield skill — that extracts the implicit flow from an existing
+`SKILL.md`, asks you which language you want, has the model write the
+program, and completes only when the generated skill passes its own
+fixture run. A conversion that was never executed is never "done".
+
 ## Try it
 
 ```

labs/22-yield/yield/docs/locus-converter.md

@@ -0,0 +1,39 @@
+# Locus derivation — the skill converter
+
+`examples/convert-skill` turns an existing prose `SKILL.md` into a Yield
+program in the operator's chosen language (Go / TypeScript / Python /
+Rust). The converter is itself a Yield skill — the pipeline that makes
+skills reliable is the pipeline that converts them.
+
+## Verdicts (models in `docs/locus/`)
+
+| model | operator | verdict |
+|---|---|---|
+| `convert-verified.json` | control.supervisory-rw | **controllable**, no violations — shipping an unverified conversion is preventable |
+| `convert-verified.json` | control.nonblockingness | **nonblocking** — every conversion ends at `SHIPPED_VERIFIED` or an honest `REPORTED_BLOCKED` |
+| `convert-verified.json` vs `convert-transcribed.json` | verification.safety-reachability (rival designs, `drv-…`) | **decided** — 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 decided comparison is the design's spine: **a conversion that was
+never executed is never "done".** `complete_untested` stays in the
+alphabet with no transition — the program's `Require(test.exit_code == 0)`
+is the refusing mechanism, and the two-attempt repair loop ends in an
+honest `Blocked`, keeping the pipeline nonblocking.
+
+## Division of labor
+
+| code owns | model owns |
+|---|---|
+| pipeline order (read → extract → choose → generate → verify) | reading the prose and extracting the implicit flow |
+| the language menu (`ask_user`, closed set) | writing the program, thin SKILL.md, runner manifest, fixtures |
+| the repair bound (≤ 2 attempts) | repairing a failing generation |
+| the evidence gate (`yskill test` exit code, observed by the supervisor) | — |
+
+## Verified end-to-end without a model
+
+`fixtures/responses.json` scripts a conversion whose destination is an
+existing valid skill, so `yskill test examples/convert-skill` exercises
+the full machinery — including the **nested** `yskill test` of the
+"generated" skill (`${YSKILL:-yskill}` lets harnesses pin the binary).
+What the scripted run cannot exercise is the model actually writing good
+code; that is exactly the part the evidence gate exists to check at live
+time.

labs/22-yield/yield/docs/locus/convert-transcribed.json

@@ -0,0 +1,161 @@
+{
+  "schema_version": 1,
+  "id": "convert-skill-transcribed-v1",
+  "subject": "Yield skill converter pipeline: extract flow from prose SKILL.md, operator picks the target language, the model generates the program, and the converter completes on the model's word that the code is correct.",
+  "evidence": [
+    {
+      "path": "labs/22-yield/yield/cmd/yskill/main.go",
+      "note": "design B: no yskill test run; the generated program is returned as text and shipped on transcription"
+    }
+  ],
+  "states": [
+    {
+      "id": "SOURCED"
+    },
+    {
+      "id": "FLOW_EXTRACTED"
+    },
+    {
+      "id": "LANG_CHOSEN"
+    },
+    {
+      "id": "GENERATED"
+    },
+    {
+      "id": "TEST_RUN"
+    },
+    {
+      "id": "TEST_PASSED"
+    },
+    {
+      "id": "TEST_FAILED"
+    },
+    {
+      "id": "SHIPPED_VERIFIED",
+      "marked": true
+    },
+    {
+      "id": "SHIPPED_UNVERIFIED"
+    },
+    {
+      "id": "REPORTED_BLOCKED",
+      "marked": true
+    }
+  ],
+  "events": [
+    {
+      "id": "read_prose",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "extract_flow",
+      "controllable": false,
+      "observable": true
+    },
+    {
+      "id": "pick_language",
+      "controllable": false,
+      "observable": true
+    },
+    {
+      "id": "generate_program",
+      "controllable": false,
+      "observable": true
+    },
+    {
+      "id": "run_generated_test",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "test_passes",
+      "controllable": false,
+      "observable": true
+    },
+    {
+      "id": "test_fails",
+      "controllable": false,
+      "observable": true
+    },
+    {
+      "id": "regenerate",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "complete_with_evidence",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "complete_untested",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "declare_blocked",
+      "controllable": true,
+      "observable": true
+    },
+    {
+      "id": "archive",
+      "controllable": false,
+      "observable": true
+    }
+  ],
+  "transitions": [
+    {
+      "from": "SOURCED",
+      "event": "extract_flow",
+      "to": "FLOW_EXTRACTED",
+      "evidence": [
+        0
+      ]
+    },
+    {
+      "from": "FLOW_EXTRACTED",
+      "event": "pick_language",
+      "to": "LANG_CHOSEN",
+      "evidence": [
+        0
+      ]
+    },
+    {
+      "from": "LANG_CHOSEN",
+      "event": "generate_program",
+      "to": "GENERATED",
+      "evidence": [
+        0
+      ]
+    },
+    {
+      "from": "GENERATED",
+      "event": "complete_untested",
+      "to": "SHIPPED_UNVERIFIED",
+      "evidence": [
+        0
+      ]
+    },
+    {
+      "from": "SHIPPED_UNVERIFIED",
+      "event": "archive",
+      "to": "SHIPPED_UNVERIFIED",
+      "evidence": [
+        0
+      ]
+    }
+  ],
+  "spec": {
+    "description": "A converted skill must never ship without its fixture run passing under the real supervisor.",
+    "forbidden_states": [
+      "SHIPPED_UNVERIFIED"
+    ]
+  },
+  "targets": {
+    "selector": "marked"
+  },
+  "unknowns": [
+    "Design B ships whatever the model asserts compiles; nothing observes the generated program executing."
+  ]
+}

labs/22-yield/yield/docs/locus/convert-verified.json

@@ -0,0 +1,57 @@
+{
+  "schema_version": 1,
+  "id": "convert-skill-verified-v1",
+  "subject": "Yield skill converter pipeline: extract flow from prose SKILL.md, operator picks the target language, the model generates the program, and the converter completes only after the generated skill passes its own fixture run under yskill test.",
+  "evidence": [
+    { "path": "labs/22-yield/yield/cmd/yskill/main.go", "note": "yskill test: scripted fixture run of a skill directory; exit code is observed fact" },
+    { "path": "labs/22-yield/yield/internal/guard/guard.go", "note": "evidence-bound completion: a failed requirement prevents run.completed" }
+  ],
+  "states": [
+    { "id": "SOURCED" },
+    { "id": "FLOW_EXTRACTED" },
+    { "id": "LANG_CHOSEN" },
+    { "id": "GENERATED" },
+    { "id": "TEST_RUN" },
+    { "id": "TEST_PASSED" },
+    { "id": "TEST_FAILED" },
+    { "id": "SHIPPED_VERIFIED", "marked": true },
+    { "id": "SHIPPED_UNVERIFIED" },
+    { "id": "REPORTED_BLOCKED", "marked": true }
+  ],
+  "events": [
+    { "id": "read_prose", "controllable": true, "observable": true },
+    { "id": "extract_flow", "controllable": false, "observable": true },
+    { "id": "pick_language", "controllable": false, "observable": true },
+    { "id": "generate_program", "controllable": false, "observable": true },
+    { "id": "run_generated_test", "controllable": true, "observable": true },
+    { "id": "test_passes", "controllable": false, "observable": true },
+    { "id": "test_fails", "controllable": false, "observable": true },
+    { "id": "regenerate", "controllable": true, "observable": true },
+    { "id": "complete_with_evidence", "controllable": true, "observable": true },
+    { "id": "complete_untested", "controllable": true, "observable": true },
+    { "id": "declare_blocked", "controllable": true, "observable": true },
+    { "id": "archive", "controllable": false, "observable": true }
+  ],
+  "transitions": [
+    { "from": "SOURCED", "event": "extract_flow", "to": "FLOW_EXTRACTED", "evidence": [0] },
+    { "from": "FLOW_EXTRACTED", "event": "pick_language", "to": "LANG_CHOSEN", "evidence": [0] },
+    { "from": "LANG_CHOSEN", "event": "generate_program", "to": "GENERATED", "evidence": [0] },
+    { "from": "GENERATED", "event": "run_generated_test", "to": "TEST_RUN", "evidence": [0] },
+    { "from": "TEST_RUN", "event": "test_passes", "to": "TEST_PASSED", "evidence": [0] },
+    { "from": "TEST_RUN", "event": "test_fails", "to": "TEST_FAILED", "evidence": [0] },
+    { "from": "TEST_FAILED", "event": "regenerate", "to": "GENERATED", "evidence": [0] },
+    { "from": "TEST_FAILED", "event": "declare_blocked", "to": "REPORTED_BLOCKED", "evidence": [1] },
+    { "from": "TEST_PASSED", "event": "complete_with_evidence", "to": "SHIPPED_VERIFIED", "evidence": [1] },
+    { "from": "SHIPPED_VERIFIED", "event": "archive", "to": "SHIPPED_VERIFIED", "evidence": [0] },
+    { "from": "REPORTED_BLOCKED", "event": "archive", "to": "REPORTED_BLOCKED", "evidence": [0] }
+  ],
+  "spec": {
+    "description": "A converted skill must never ship without its fixture run passing under the real supervisor.",
+    "forbidden_states": ["SHIPPED_UNVERIFIED"]
+  },
+  "targets": { "selector": "marked" },
+  "unknowns": [
+    "Design A refuses complete_untested structurally: the event is in the alphabet, but the converter program offers no transition — Require(test.exit_code == 0) stands between GENERATED and completion.",
+    "Regeneration is bounded by a retry counter in the converter program; the model treats the bound as data, not modeled state."
+  ]
+}