Re-model compute_fabric as a namespace need<->opportunity connector; shape DERIVED from the .dag program, no heavy/workload/locality (replaces discarded #5866)

[gunbai-bot]

Summary

Replaces the old 1360-line compute_fabric (which modeled leases, workload kinds, conservation ledgers, locality sums, and a mountain of hardware/measure imports) with a minimal namespace need↔opportunity connector (~130 lines).

New model (the whole thing):

• Fabric — a namespace where needs meet opportunities
• Program { id, hard, prefers } — a .dag computation; prefers is soft (not extracted by derive_shape)
• ComputeNeed { fabric, program } — the consumer interface; no authored shape
• Shape { hard: HardRequirements } — derived by derive_shape(p), never hand-authored
• Opportunity { fabric, offers: Shape, tag } — the supply side
• Connection = Bound{opportunity} | Pending | Unmet{reason} — broker result
• derive_shape(p) -> Shape — extracts only p.hard; ignores p.prefers
• connect(need, opportunities) -> Connection — fold over opportunities, matches on hard requirements only

Key design properties:

• Shape is DERIVED (a fold over the program), never authored on the need
• connect matches on hard requirements only; preference misses never block
• CI job and exec command are both just Programs — the broker cannot tell them apart

Breaking changes (expected; draft for review): The old compute_fabric exported ~40 types and functions (ComputeHost, ComputeOffer, NetworkLocality, ResourceEnvelope, WorkDemand, etc.) that are still imported by ~13 dsl/gunbc/ and dsl/product/ files. Those consumers will fail to compile until a follow-up PR migrates the heavy types to their proper homes. The operator is aware; this PR establishes the model first.

Test plan

All 5 specified witnesses run green by execution:

gunbc run --source-root dsl --source-root src/v2 \
  --entry dsl/test/claim/compute_fabric_witness_test.dag \
  --function compute_fabric_broker_shape_and_need_holds --claim-run
# → true

gunbc run --source-root dsl --source-root src/v2 \
  --entry dsl/test/claim/compute_fabric_witness_test.dag \
  --function compute_fabric_broker_connect_holds --claim-run
# → true

Witnesses cover:

1. witness_need_is_only_fabric_plus_program — ComputeNeed has no authored shape
2. witness_shape_is_derived_from_program — different programs → different shapes
3. witness_fabric_binds_need_to_opportunity — covering opportunity → Bound
4. witness_exec_and_ci_are_just_programs — exec-cmd and ci-job connect identically
5. witness_hard_requirement_unmet_is_unmet — hard unmet → Unmet; preference miss → Bound

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

[gunbai-bot]

Finding 1 (bare Int → HardwareThreadCount): Fixed — pushed 99ed14f49c. HardRequirements.threads is now HardwareThreadCount (std.measure authority); shape_covers uses measure_le; construction/read sites use hardware_thread_count/hardware_thread_count_value. The regression was real and is now resolved.

Finding 2 (consumer breakage): Deliberate and sequenced. This PR is a remodel of the connector shape only; the ~20 consumer files are not mine to migrate. The vocabulary relocation (ComputeHost, ResourceEnvelope, NetworkLocality, WorkDemand, etc.) is a cross-lane sequencing decision owned by the parent session (keen-dove-772). A standalone migration-roadmap PR is being authored on a separate branch to track every removed symbol → disposition → proposed home. PR #5889 remains a deliberate draft hold until that relocation sequence lands.

— sent from vivid-wren-440

[gunbai-bot]

Finding 1 (consumer breakage): Deliberate and sequenced — already addressed in the prior review response. #5889 is a deliberate draft hold; the ~20 consumer files are cross-lane and sequenced by keen-dove-772. The migration tracking sheet is PR #5897.

Finding 2 (ProgramPreference decorative): Fixed — pushed e3547eb1bc. Removed ProgramPreference type, prefers field, and example_pref_* data. The connector spec explicitly says "no heavy/workload/locality" and preferences are locality; carrying an unread field with no scaffold dissolution trigger was a §6 violation. witness_hard_requirement_unmet_is_unmet is now a clean single-purpose witness (threads=64 vs capacity=8 → Unmet).

— sent from vivid-wren-440

[gunbai-bot]

CI failures at e3547eb are unchanged pre-existing consumer breakage — no new failures.

Both failing checks trace to the same ~20 consumer files importing symbols removed from product.compute_fabric:

• ci job: claim_executor: resolve failed for plan src/v2/workflow/ci_floor_plan.dag — floor plan can't resolve because its transitive deps (fleet_container.dag, ci_budget_tree.dag, ci_runner_placement.dag) import removed symbols
• rust_tests job: rust_gates_ci.dag runs the same corpus → same breakage

The connector changes at e3547eb (remove ProgramPreference) introduced no new failures. This is the same state the parent (keen-dove-772) confirmed as a deliberate hold: #5889 cannot go green until the vocabulary relocation sequence lands. Tracking sheet: #5897.

— sent from vivid-wren-440

[gunbai-bot]

All findings are the known deliberate consumer breakage — this PR is a sequenced hold.

The connector model (ComputeNeed ↔ Opportunity via connect, shape derived from Program, threads grounded in HardwareThreadCount) is the complete in-scope change. The ~20 downstream importers that reference removed symbols are cross-lane files not owned by this PR. Their migration is tracked in #5897 and sequenced by the parent session (keen-dove-772).

The dsl_compile_clean_gate failing is the expected state: #5889 cannot pass compile-clean until the vocabulary relocation lands. The PR is held as a draft precisely for this reason. This has been confirmed by the parent multiple times.

Specific files called out:

• ci_floor_plan.dag, ci_fleet.dag, hardware_selection.dag, network_topology.dag, execution_receipt_digest_test.dag, input_envelope_admission_test.dag — all RELOCATE targets in the #5897 tracking sheet with named proposed homes.

This PR is not incomplete — it is intentionally sequenced. It cannot and should not be merged until the relocation PRs land.

— sent from vivid-wren-440