Floor execution requirement moves to gunbc.floor_demand (floor_demand witness closure 1,761 -> 52)

[gunbai-bot]

Cut (B) of the floor_demand closure item (jolly-boar-500, from sleek-ibex-207). It is independent of cut (A), #12783. Both are module-boundary moves: no resolver change, no new concept, no re-export shim.

Derivation

• gunbc.floor_demand closes over about 47 modules.
• Its own witness, and several production consumers, read the floor's execution requirement (floor_execution_requirements). That requirement lived in gunbc.fabric_witness_run, whose subject is building the witness run's Work and CI job, and which closes over about 1,255 modules on main.
• The requirement's memory is floor_demand's own floor_memory_requirement(). The rest (thread count, capability, trust domain, isolation profile) completes what the floor asks of a host.
• So it is a fact about floor demand, placed in the module that consumes it most heavily.

Change

• Moved to gunbc.floor_demand, with names unchanged:
  • floor_execution_requirements
  • floor_isolation_requirement
  • gunbc_internal_fleet_trust_domain
• The trust domain moves with them. It has no shared home: trust domains are declared beside their consumers (spark_serving_trust_domain lives in gunbc.serving.turn_admission). Its role here is the domain floor Work is declared under, which runner_attempt_launch admits. Leaving it in fabric_witness_run would create a cycle, and a new module for one constant would be a new authority.
• floor_demand gains imports of product.fabric.{work, isolation} and ResourceEnvelope. Nothing under dag/product imports gunbc.*, so there is no cycle.

Every importer of the moved symbols (enumerated, since not all are under the required gate)

Now imported from gunbc.floor_demand:

• gunbc.fabric_witness_run: keeps only floor_execution_requirements, which it uses itself.
• gunbc.fabric_floor_dispatch
• gunbc.fabric_control_plane: previously imported it through fabric_floor_dispatch, a re-export; now imports it directly from its home.
• gunbc.runner_attempt_launch: import, plus the comment citing the trust domain's module.
• gunbc.runner_microvm
• gunbc.runner_microvm_shakedown
• test.claim.floor_demand_witness
• test.claim.fabric_isolation_witness
• test.claim.runner_microvm_slot_controller_witness
• test.claim.runner_throughput_qualification_witness: import, plus its decl_ref(module_path: "gunbc.fabric_witness_run", decl_name: "floor_execution_requirements"), which now names gunbc.floor_demand.

No Rust references the moved symbols.

Why the floor_demand witness keeps reading floor_execution_requirements

The suggestion was that it read floor_memory_requirement() directly. Its claim the_work_memory_requirement_is_the_held_set_peak_or_absent checks that the held-set standing reaches the Work's execution requirement (floor_execution_requirements().shape.envelope.memory). That is an inhabitance claim on the route into Work (DESIGN §3), and reading the upstream value would weaken it. After the move it imports the requirement from gunbc.floor_demand, the module under test, so the closure concern is gone without losing the route.

Evidence

Each affected witness was run with gunbc run --claim-run, one cold process per file, 16 GiB memory.max leaf, on current main plus this change:

witness	result	closure files	time	peak RSS
floor/floor_demand_witness_test	29 PASS	52 (main: 1,761)	54s (main: 251s)	7.5 GB (main: 15.2 GB)
fabric/fabric_isolation_witness_test	6 PASS	51	49s	7.5 GB
runner/runner_microvm_slot_controller_witness_test	13 PASS	1,805	253s	15.9 GB
runner/runner_throughput_qualification_witness_test	28 PASS	1,782	238s	15.6 GB
fabric/fabric_floor_dispatch_witness_test	4 PASS	1,763	228s	15.3 GB
fabric/fabric_witness_run_test	refused at resolve	—	—	—

• fabric_witness_run_test refuses on main too, as a standalone run, with the same class: effect summary incomplete … names no registry row. On main the named callees include product.fabric.isolation.current_runner_slot_profile and product.fabric.supply.offer_affordability_for. The class predates this PR and it does not introduce it. I'm noting it rather than fixing it here.
• The runner and dispatch witnesses keep large closures through their other imports. Cut (A), Fleet build tuning to gunbc.fleet_workflow_steps; emitter drops unused heavy import (floor_demand closure 1,761 -> 218) #12783, shrinks the part of that which comes from fabric_witness_run.

Do not merge; landing goes through the operator.

🤖 Generated with Claude Code

[gunbai-bot]

Addressing review 73251: fixed in d71833a. My move cut each declaration's body and left its leading // rationale behind in fabric_witness_run.dag, where it read as one long comment on the SOURCE ROOTS declaration. All four blocks now sit directly above their own declarations in gunbc.floor_demand:

• "WHAT A FLOOR RUN REQUIRES OF ITS EXECUTOR…" (with its CONSUMPTION STATUS / NEXT-RUNG TRIGGER) → floor_isolation_requirement
• "THE FLEET'S OWN TRUST DOMAIN, AS A ROW…" → gunbc_internal_fleet_trust_domain
• "OUR FLOOR IS arm64. THE CLASS IS NOT." (its "this function previously spelled the architecture inside a capability tag" refers to the capabilities field) plus "THE FLOOR IS CONTROL-PLANE WORK…" (the envelope reasoning with the storage-frontier TRIGGER) → floor_execution_requirements

fabric_witness_run.dag no longer carries any of them. The relocation note is now part of the leading block rather than a standalone block (§4c admits only attached leading blocks). floor_demand_witness_test passes 29/29 and fabric_isolation_witness_test 6/6 on this head.

— sent from bold-bat-516