docs(device-mesh): establish upstream port authority

[fivetide]

Summary

Non-blocking governance proposal: G1-G5 implementation may proceed from current upstream while referencing the reviewed tracker on the public fivetide/hipfire branch. Merging this PR would add upstream CI enforcement and atomic status updates, but is not a runtime or architecture prerequisite.
Establish the machine-checkable upstream-native authority for replacing superseded PR #527, following the approved grouped port in #666.

This PR contains no runtime, model-admission, GPU, or generation behavior change. It defines:

• 68 fine-grained domain obligations with historical evidence disposition;
• 16 grouped delivery change sets (G0–G15);
• 22 executable seam gates;
• 5 topology-class evidence campaigns and final closure packet;
• exact source/owner/route/evidence/revert contracts;
• fail-closed status promotion rules;
• the docs/INDEX.md authority link;
• a stdlib checker and focused malformed-tracker tests.

The grouped approach keeps evidence precision without requiring one PR per obligation. Reset/lifecycle, manifest/loading, topology/admission, and the interface-safe MoE Step lane can develop in parallel; merge ordering follows the seams each change set consumes.

Closes the planning portion of #666 G0. PR #527 remains open as historical forensic provenance and is explicitly marked superseded.

Architecture contracts

• LoadedModel / ArchModel remains the sole model-state and teardown owner.
• hipfire-generate plus engine scheduler/terminal remains the sole generation/lifecycle authority.
• Hardware ownership must move by clean cutover with no Gpus compatibility re-export.
• Raw PP/TP/EP requests are normalized once at admission; placement consumes admitted topology only.
• Step collective scheduling depends on DeviceMesh plus the pure manifest schedule, not WeightStore representation.
• Every positive route carries an exact fail-closed/rollback route and side-effect assertions.
• Historical, semantics-only, emulated, failed, host-local, or retired-gate evidence cannot promote current upstream status.
• Physical completion requires pinned model/prompt-or-image/binary/GPU/topology/ROCm/RCCL identities and per-obligation pass mapping.

Commits

1. 4429d7df3dff048522e55c2c0d6081b40bd45976 — tracker/checker/tests/index implementation
2. 514fd489326b70d1d796849d19929c8e85c76b5d — evidence metadata pinned to commit 1

The two-commit shape avoids self-referential evidence hashes: executable checker/test evidence is pinned to the exact implementation commit; the second commit records that evidence only.

Verification

Clean upstream baseline before changes:

• cargo build — pass
• cargo test — 3,410 passed, 12 ignored, 0 failed

PR-focused checks:

• python3 -m py_compile scripts/check-device-mesh-port-tracker.py tests/test_device_mesh_port_tracker.py — pass
• python3 scripts/check-device-mesh-port-tracker.py — pass (68 obligations, 16 groups, 5 campaigns)
• pytest -q tests/test_device_mesh_port_tracker.py — 36 passed
• git diff --check — pass

Independent final review: spec PASS / quality APPROVED / no findings.

scripts/check-env-docs.py remains red on pre-existing HIPFIRE_* ownership drift in untouched runtime sources. This PR introduces no environment variable or runtime ownership changes; the focused authority/schema/index checks above are the binding route.

Scope guard

No change to:

• docs/VALIDATION.md authority;
• docs/admissions.yml authority;
• runtime, loader, dispatcher, architecture, generation, or kernel code;
• model/route admission;
• PR [Superseded] refactor: unify device mesh, model parallelism, and generation dispatch #527 branch history.

Follow-up

After merge, #666 schedules the grouped upstream-native lanes:

• G1 topology
• G2 admission
• G3 manifest/loading
• G4 reset/lifecycle
• G5 MoE Step architecture
• G6–G11 family cutovers
• G12 generation/batching
• G13–G15 axes
• physical evidence campaigns and aggregate closure

[fivetide]

Implementation PR chain from this tracker:

• feat(hardware): establish named device mesh ownership #673 G1 topology — ready, root
• feat(loader): centralize source-aware parallel admission #674 G2 admission — draft, depends on feat(hardware): establish named device mesh ownership #673
• feat(runtime): add device-mesh weight manifest and store pilot #675 G3 manifest/store — draft, depends on feat(hardware): establish named device mesh ownership #673
• fix(runtime): consolidate generation lifecycle ownership #676 G4 lifecycle — ready, independent
• feat(dispatch): seal collective-aware MoE step execution #677 G5 MoE steps — draft, depends on feat(runtime): add device-mesh weight manifest and store pilot #675/feat(hardware): establish named device mesh ownership #673; feat(loader): centralize source-aware parallel admission #674 required before complete production multi-device enablement

Canonical dependency and landing-order note: #666 (comment)

[Kaden-Schutt]

Note for anyone landing here: this PR shows MERGED, but its content was backed out of master the same evening in a0fca0d6d after the G2 admission layer refused every Qwen3.5-family registry model on hardware. The merged state is preserved at tag archive/device-mesh-merged-20260901 (541b33c33). Full account, both reproductions, and the re-land path: #666 (comment)