Record accepted SDK preparation and bind OpenHands fact workspace

blueprints/convergence-practice/clean-resolution-20261002/open-work.json

@@ -3,14 +3,15 @@
   "recorded_date_utc": "2026-10-02",
   "goal_status": "active",
   "base_revision": "18eea2c1de992b46c266d79ef0cc40f93c9fb943",
+  "status_assessed_at_revision": "73ca1e8b53383f5bd5e1a59a06eade50a0086370",
   "scope": "Full user-requested foundation resolution serving the north star; the SDK oracle is one assigned unit, not a replacement objective.",
   "requirements": [
     {
       "id": "github-practice",
       "state": "partially_verified_followups_open",
-      "evidence": "Read-only remote audit found eight app-bound required checks, full-SHA action policy, restricted publishers and no ruleset bypass at the observed main.",
-      "remaining": ["Qualify the gh 2.102.0 security update through the installation owner", "Repair owned failing PR checks and require complete current-head coverage", "Review catalog-bot metadata repair in PR611 and observe its next authorized hosted proposal"],
-      "sources": ["https://github.com/cli/cli/releases/tag/v2.102.0", "docs/decisions/2026-10-02-github-automation-practice.md"]
+      "evidence": "Read-only remote audit found eight app-bound required checks, full-SHA action policy, restricted publishers and no ruleset bypass at the observed main. Isolated legacy-host qualification of gh 2.102.0 passed the unchanged upstream race/integration command (250 packages: 245 native cached results and five fresh executions; 60 packages without tests), native build and local binary smoke. Original failed conditions and source/output bindings were retained and independently reviewed. Catalog metadata source review is accepted; hosted behavior and current-head publication gates remain separate.",
+      "remaining": ["Integrate the gh 2.102.0 source qualification through the installation owner without inferring host adoption", "Require complete current-head coverage before landing the catalog metadata fix", "Observe the next authorized hosted catalog proposal and its source/lane metadata"],
+      "sources": ["https://github.com/cli/cli/releases/tag/v2.102.0", "docs/decisions/2026-10-02-github-automation-practice.md", "https://github.com/seathatflowsinourveins/native-agent-stack/pull/607#issuecomment-5959705875", "https://github.com/seathatflowsinourveins/native-agent-stack/pull/611#issuecomment-5959596846"]
     },
     {
       "id": "maintained-layer-selections",
@@ -28,10 +29,10 @@
     },
     {
       "id": "sdk-workers-researchers-reviewers",
-      "state": "oracle_and_architecture_reviewed_final_binding_and_native_execution_open",
-      "evidence": "Three local oracle-discrimination runs retained with source and result hashes; target SDK pin identities checked. Scoped Claude and Astra reviews at 6537a17a1 accept the oracle and cross-runtime architecture; the requested evidence-wording correction is incorporated for final binding review.",
-      "remaining": ["Final evidence/head binding review", "Actual selected SDK tool task and result/usage capture", "Separately qualify resume, recovery, selected MCP integration and role boundaries"],
-      "sources": ["examples/native-sdk-acceptance/README.md", "blueprints/convergence-practice/clean-resolution-20261002/oracle-experiment.json"]
+      "state": "source_preparation_accepted_and_published_native_execution_open",
+      "evidence": "Three local oracle-discrimination runs retained with source and result hashes; target SDK pin identities checked. Scoped Claude and Astra reviews accepted the oracle and cross-runtime preparation. Final source/evidence binding accepted 45ffc7d4ec0e5063278299ad736c36c497732ec3 with all seven frozen artifacts unchanged; publication merged as 73ca1e8b53383f5bd5e1a59a06eade50a0086370 after all eight required checks passed. This is source preparation, not target SDK execution.",
+      "remaining": ["Builder's selected readiness and safe ownership handoff", "Actual selected SDK tool task with run-specific source/route binding and result/usage capture", "Separately qualify resume, recovery, selected MCP integration and role boundaries"],
+      "sources": ["examples/native-sdk-acceptance/README.md", "blueprints/convergence-practice/clean-resolution-20261002/oracle-experiment.json", "https://github.com/seathatflowsinourveins/native-agent-stack/pull/609#issuecomment-5958803761", "https://github.com/seathatflowsinourveins/native-agent-stack/pull/609#issuecomment-5959596671"]
     },
     {
       "id": "token-skills-orchestration",

examples/native-sdk-acceptance/README.md

@@ -1,6 +1,6 @@
 # Selected SDK useful-work acceptance preparation
 
-Status: source/oracle preparation; target execution and architecture acceptance pending.
+Status: source/oracle preparation accepted and published; target execution remains pending.
 North-star action served: qualify the runtime workers that perform bounded research and engineering tasks before relying on their work in the US-equities research stack.
 
 The existing [selected acceptance plan](https://github.com/seathatflowsinourveins/native-agent-stack/blob/fc303db44d52c0d08f9d568866bec76c411fd3aa/evidence/artifacts/new-wsl-install-plan-20261002/accept.sh) runs supported interfaces but discards application stdout at line 52. Its stage status cannot prove the answer, produced file, tool outcome or native usage. This preparation supplies explicit useful-work and evidence requirements alongside those interfaces. The existing runtimes remain responsible for execution.
@@ -29,10 +29,18 @@ Keep every failed, timed-out or skipped attempt and the reason. Preserve origina
 
 The upstream [hello-world example](https://github.com/OpenHands/software-agent-sdk/blob/1e1390acc8788346ba4804c34323284009bf3f5e/examples/01_standalone_sdk/01_hello_world.py) asks for three facts about the current project and writes `FACTS.txt`. Freeze the assigned repository/revision and a source allow-list before execution. The independent verifier must read that exact resulting file, identify three distinct factual claims and attach a source path, content hash and supporting location to each. Reject missing/empty output, unsupported claims and facts from a different repository. Record extra claims and their correctness rather than silently dropping them.
 
+The [pinned example](https://github.com/OpenHands/software-agent-sdk/blob/1e1390acc8788346ba4804c34323284009bf3f5e/examples/01_standalone_sdk/01_hello_world.py#L24) uses `os.getcwd()` as its workspace. The [selected install plan](https://github.com/seathatflowsinourveins/native-agent-stack/blob/b5079894cb08508b09214613dac5774e0fb017ef/evidence/artifacts/new-wsl-install-plan-20261002/install-plan.json#L2314) already assigns the OpenHands v1.50.1 source checkout at `$tool_root/openhands-source`; its [acceptance caller](https://github.com/seathatflowsinourveins/native-agent-stack/blob/b5079894cb08508b09214613dac5774e0fb017ef/evidence/artifacts/new-wsl-install-plan-20261002/accept.sh#L537) launches the unchanged example from that directory. Before execution, the builder handoff must verify the actual target HEAD and source bytes against `1e1390acc8788346ba4804c34323284009bf3f5e`, record the resolved launch directory and `FACTS.txt` destination, and retain the resulting file for independent grading. A new project selection is unnecessary.
+
+For this local oracle, the source allow-list is the complete versioned tree at that commit: root [Git tree `64e12e9ee91c3ec48e9f7c516ca4a8c178539d4d`](https://api.github.com/repos/OpenHands/software-agent-sdk/git/trees/64e12e9ee91c3ec48e9f7c516ca4a8c178539d4d). This records the admissible evidence corpus before execution and preserves freedom to return any three supported project facts. Retain a source path, content hash and supporting location for each graded claim against this immutable corpus. Filesystem permissions, runtime sandboxing and the actual target binding require separate verification.
+
+Keep the supporting upstream test scopes distinct. The [example harness](https://github.com/OpenHands/software-agent-sdk/blob/1e1390acc8788346ba4804c34323284009bf3f5e/tests/examples/test_examples.py#L50) excludes this hello-world example. The [mock cross-test](https://github.com/OpenHands/software-agent-sdk/blob/1e1390acc8788346ba4804c34323284009bf3f5e/tests/cross/test_hello_world.py#L186) asks for a `hello.py` file despite its nearby comment saying it sends the same message; it does not grade project facts. The [metrics example](https://github.com/OpenHands/software-agent-sdk/blob/1e1390acc8788346ba4804c34323284009bf3f5e/examples/01_standalone_sdk/13_get_llm_metrics.py#L65) changes the task, adds MCP fetching and deletes its facts file. Reuse its supported metrics interfaces only through the declared capture adaptation, preserving the selected hello-world task and its output.
+
 Preselecting three arbitrary facts and requiring the unchanged open-ended prompt to return those exact facts would change its semantic contract. This rubric permits any three supported project facts; the actual source allow-list and grading result must be frozen/retained for the selected run. An independent reviewer must accept this rubric before target execution. It is a local semantic integration check, not an upstream test or comparative benchmark.
 
 ## Review and execution gates
 
-The consequential cross-runtime acceptance review was assigned to the required Astra/Max role. That agent failed at the account usage limit before returning a judgment; no acceptance is inferred. The assigned fixture worker also failed before writing files. The coordinator prepared the source fixture and local checks; a separate reviewer must inspect them. The [resolution envelope](../../blueprints/convergence-practice/clean-resolution-20261002/README.md) retains the broader open work.
+The initial Astra/Max review and assigned fixture worker failed at the account usage limit before returning a judgment or writing files. Those failed attempts remain in the [resolution envelope](../../blueprints/convergence-practice/clean-resolution-20261002/README.md). The coordinator prepared the fixture and local checks; subsequent [independent Claude review](https://github.com/seathatflowsinourveins/native-agent-stack/pull/609#issuecomment-5958056799) and [Astra/Max source/evidence acceptance](https://github.com/seathatflowsinourveins/native-agent-stack/pull/609#issuecomment-5958803761) closed the preparation review. The final binding is to `45ffc7d4ec0e5063278299ad736c36c497732ec3`, with all seven frozen fixture/oracle/control/output artifacts unchanged. [Publication](https://github.com/seathatflowsinourveins/native-agent-stack/pull/609#issuecomment-5959596671) merged that preparation as `73ca1e8b53383f5bd5e1a59a06eade50a0086370` after all eight required checks passed.
+
+These acceptances cover source preparation. A selected target readiness handoff, run-specific source and route binding, actual SDK task output/usage, independent task grading and separate lifecycle/role qualification remain open. The [full requirement map](../../blueprints/convergence-practice/clean-resolution-20261002/open-work.json) retains those gates; publication does not close the sustained resolution goal.
 
 A future comparative claim must use the repository's selected upstream evaluation harness with preregistered tasks and scoring. No comparative or token-saving claim is made here, and this directory introduces no model runner.