Accept CMUX repository-owned workload profile requests

.github/workflows/ci.yml

@@ -80,6 +80,8 @@ jobs:
               Path("scripts/cmux_fleet_bootstrap.py"),
               Path("scripts/test-cmux-fleet.py"),
               Path("scripts/test-cmux-fleet-bootstrap.py"),
+              Path("scripts/cmux_workload_request.py"),
+              Path("scripts/test-cmux-workload-request.py"),
               Path("scripts/test-verify-focused.py"),
               Path("scripts/verify-changed-rustfmt"),
               Path("scripts/test-hot-run.py"),
@@ -121,6 +123,7 @@ jobs:
           python3 scripts/test-owned-linux-admission.py
           python3 scripts/test-cmux-fleet.py
           python3 scripts/test-cmux-fleet-bootstrap.py
+          python3 scripts/test-cmux-workload-request.py
           python3 scripts/test-verify.py
           python3 scripts/test-verify-changed-rustfmt.py
           python3 scripts/test-front-door.py

docs/CMUX_WORKLOAD_PROFILES.md

@@ -0,0 +1,114 @@
+# CMUX repository-owned workload profiles
+
+CMUX owns the meaning of CMUX workload profiles. Glaeda may admit, place, execute, cache, time, and recover those workloads without copying their command lines or redefining pass/fail semantics.
+
+The paired CMUX implementation is tracked by `manaflow-ai/cmux#13411`.
+
+## Request boundary
+
+`scripts/cmux_workload_request.py` accepts a bounded semantic request containing:
+
+- exact `manaflow-ai/cmux` commit and tree;
+- CMUX profile ID and explicit semantic generation;
+- benchmark state class;
+- bounded integer semantic parameters, such as an app-host shard number;
+- caller correlation and an advisory reuse hint.
+
+It rejects caller-controlled argv, shell text, cwd, host paths, backend/machine choice, resource values, cache roots, attempt identity, or cleanup authority.
+
+A planned request resolves only this fixed adapter:
+
+```json
+{
+  "kind": "cmux-repository-profile/v1",
+  "repository_runner": "scripts/ci/cmux_workload_profile.py",
+  "semantic_result_contract": "cmux-workload-result/v1"
+}
+```
+
+Glaeda does not keep a copy of the CMUX registry. Profile validity is checked after exact source materialization by the runner inside that CMUX tree.
+
+## Physical execution
+
+A physical adapter follows this sequence:
+
+1. materialize the exact requested CMUX commit/tree;
+2. choose an eligible backend and admitted cache/state class from Glaeda-owned observations;
+3. run the checked-in CMUX profile runner's `plan` command and require the same source, profile ID/generation, state class, and semantic parameters;
+4. allocate a task-private state root and any reviewed runtime inputs;
+5. run the same CMUX profile runner and retain the exact canonical `cmux-workload-result/v1`;
+6. wrap its digest and terminal state in Glaeda's physical receipt.
+
+The generic repository adapter may translate the semantic request into the fixed CMUX runner interface. It does not reconstruct Xcode, test-selector, package, release, or developer-build commands.
+
+Runtime-input paths stay physical. Their content identities are published by the CMUX semantic result when the profile contract requires them.
+
+## Identity
+
+Three identities remain separate:
+
+- caller correlation: external request/work references;
+- semantic execution binding: exact source + CMUX profile/generation + state class + semantic parameters + fixed adapter generation;
+- physical attempt: Glaeda backend, node, leases, cache placement, resources, timing, cleanup, and recovery state.
+
+Caller correlation and reuse hints do not mint another execution binding. A profile generation change does.
+
+## Result correlation
+
+`observe` accepts only canonical `cmux-workload-result/v1` bytes and verifies:
+
+- exact repository/commit/tree;
+- exact profile ID/generation;
+- exact semantic parameters;
+- requested benchmark state class;
+- presence of the CMUX semantic validator, artifact collection, and cleanup evidence.
+
+It validates the shared `cmux-workload-result/v1` integrity envelope before projection: the result schema is closed, nested identity/evidence shapes are typed and bounded by the document ceiling, the toolchain identity must match its observations, timestamps and cleanup are self-consistent, and a `passed` result must have exit zero, complete required-artifact validation, and clean process settlement. These are interface-integrity checks; Glaeda does not re-run or reinterpret CMUX's profile-specific validators.
+
+It then projects the CMUX terminal vocabulary without redefining profile semantics:
+
+- `passed` -> `succeeded`;
+- `failed` -> `failed`;
+- `timed_out` -> `timed_out`;
+- `ambiguous` -> `ambiguous`.
+
+The outer observation stores the exact CMUX result digest. Glaeda-specific machine and execution evidence belongs beside that digest in the physical receipt.
+
+## Fleet roles and routing
+
+The first CMUX mapping is:
+
+```yaml
+cmux_linux_ci:
+  acceptance:
+    profile: cmux.ci.guard
+    generation: 1
+
+cmux_macos_native_build:
+  acceptance:
+    profile: cmux.macos.dev-check
+    generation: 1
+```
+
+Role acceptance consumes the current profile identity from CMUX rather than a Glaeda-owned recipe. Routing experiments can compare backends only when the CMUX semantic comparison identity and benchmark context agree.
+
+This gives #148 a repository-owned named-profile source, #546 comparable routing evidence, #547 a stable semantic operation for optimization, and #1056 an exact acceptance workload. #1057's caller/orchestrator identity remains independent from this workload identity.
+
+## Synthetic proof
+
+`docs/experiments/cmux-workload-profile/` contains a request and its deterministic planned result. It proves the GitHub/repository-only request boundary without claiming physical execution.
+
+Run:
+
+```sh
+python3 scripts/cmux_workload_request.py plan \
+  < docs/experiments/cmux-workload-profile/request.json \
+  > /tmp/cmux-workload-plan.json
+
+cmp /tmp/cmux-workload-plan.json \
+  docs/experiments/cmux-workload-profile/plan.json
+
+python3 scripts/test-cmux-workload-request.py
+```
+
+Physical Mac acceptance remains a separate approved fleet experiment.

docs/experiments/cmux-workload-profile/README.md

@@ -0,0 +1,17 @@
+# CMUX workload profile synthetic plan
+
+This fixture proves Glaeda can accept a caller-neutral CMUX semantic workload request while leaving CMUX semantics in the CMUX repository.
+
+`request.json` freezes synthetic exact source identity plus `cmux.ci.guard@1` and the `cold` state class. `plan.json` resolves only the fixed `cmux-repository-profile/v1` adapter and grants no execution, placement, resource-override, or redispatch authority.
+
+The OIDs are deliberately synthetic; this fixture exercises the request contract without claiming that a CMUX checkout or physical worker ran.
+
+Regenerate and compare:
+
+```sh
+python3 scripts/cmux_workload_request.py plan \
+  < docs/experiments/cmux-workload-profile/request.json \
+  > /tmp/cmux-workload-plan.json
+cmp /tmp/cmux-workload-plan.json \
+  docs/experiments/cmux-workload-profile/plan.json
+```

docs/experiments/cmux-workload-profile/plan.json

@@ -0,0 +1 @@
+{"authority":{"authorizes_execution":false,"authorizes_host_selection":false,"authorizes_redispatch":false,"authorizes_resource_override":false},"benchmark_state_class":"cold","correlation":{"work_ref":"cmux:work:fixture-1"},"document_type":"glaeda-cmux-workload-plan","execution_binding_sha256":"sha256:45ac07054ec2ba65cb46d039c879b6b973c33e2fe3f2290033f09981bfe1b8ff","external_request_ref":"cmux:workload:fixture-1","parameters":{},"physical_preflight":["materialize_exact_source","cmux_runner_plan_matches_frozen_source_and_profile","cmux_runner_owns_semantic_validation"],"profile":{"generation":1,"id":"cmux.ci.guard"},"request_sha256":"sha256:2ff1cc2403bfd2226d0e610fca06d74c0de221f18a252452b6cf58cc742a5d81","resolved_adapter":{"kind":"cmux-repository-profile/v1","repository_runner":"scripts/ci/cmux_workload_profile.py","semantic_result_contract":"cmux-workload-result/v1"},"schema_version":1,"source":{"commit":"1111111111111111111111111111111111111111","repository":"manaflow-ai/cmux","tree":"2222222222222222222222222222222222222222"},"state":"planned"}

docs/experiments/cmux-workload-profile/request.json

@@ -0,0 +1 @@
+{"benchmark_state_class":"cold","correlation":{"work_ref":"cmux:work:fixture-1"},"document_type":"glaeda-cmux-workload-request","external_request_ref":"cmux:workload:fixture-1","profile":{"generation":1,"id":"cmux.ci.guard"},"reuse_hint":"no_preference","schema_version":1,"source":{"commit":"1111111111111111111111111111111111111111","repository":"manaflow-ai/cmux","tree":"2222222222222222222222222222222222222222"}}