compiler: SO-3 pipeline carrier + spectral producer, and W5.4 executable placement

[gstoner]

Integrates a large concurrent change set (developed in the shared tree by the Codex agent) with the SO-3 work on #625, reconciled to one authority and verified end to end on this box.

Stacks on #625 — that PR's branch is this one's base, so this diff is the incoming work plus the reconciliation. Merge #625 first and this cleans up to just the new material.

What lands

SO-3 — one schedule authority (queue order 3 → closed)

• Pipeline lowering now requires the digest-bound tessera.pipeline_schedule.v1 carrier and stamps its Schedule Object digest onto functions and communication ops. The scalar-plan/options reconstruction path is deleted from PipelineStageInsertion / PipelineScheduleLegality / PMPasses — precisely the duplicate authority the row named.
• The spectral producer represents fused stages as registered Graph actions, consumes infer_action_dag, binds roles/resources into the content-addressed Schedule Object, and requires that digest at Schedule→Tile instead of re-deriving a policy hash.
• MegaMoE (from compiler: SO-3 — the MegaMoE producer consumes inferred dependence edges #625) consumes inferred edges with the hand DAG demoted to a fail-closed coverage oracle.

W5.4-RESHARD-1 — executable placement (queue order 5 → closed)

• Fixed-point placement derives exact mesh-sized local result types and inserts explicit reshard SSA at the consumer boundary using registered tessera.slice local shards rather than same-shaped fake collectives. Plan digest, subgroup, matching rounds, and the nested-region path survive Graph→Schedule→Tile; a deterministic mock-mesh executor consumes that SSA directly.
• Stated plainly in the queues: mock execution is not RCCL/multi-process evidence and cannot satisfy DIST-NATIVE-1.

Reconciliation done during the merge

• One S5 authority. The incoming set added NUMERIC-POLICY-CARRIER-1 at a new Order 6 and renumbered Orders 5–14, colliding with NUMPOL-CARRIER-1 (queue row 3b) already merged in docs: NUMPOL-CARRIER-1 — the S5 numeric_policy carrier gets an owning queue row #624. The merged row is kept; the duplicate row and the renumbering are dropped; the four backend entries that referenced the new name now point at row 3b. Orders 4–14 keep the numbers the plan and audit prose reference.
• Both sync keys preserved in all four backend queues: SCHEDULE-AUTHORITY-RESHARD-2026-08-24 (this work) and SO3-INFER-EDGES-2026-08-24 (compiler: SO-3 — the MegaMoE producer consumes inferred dependence edges #625's inference correction).

Two defects found while verifying

1. mypy ratchet break in the new mock-mesh executor (a ufunc-typed binding rebound to a lambda table) — fixed.
2. check-tessera-ir was RED on main for everyone. tests/tessera-ir/phase2/x86_composed_layout_exec.mlir, added by the merged layout work, has no RUN: line and is fixture data owned by tests/unit/test_x86_composed_layout_exec.py; lit reported it Unresolved and failed the target. Marked UNSUPPORTED: true with the reason — its owning pytest still passes, and the suite is green again.

Evidence (Strix Halo)

C++ ninja -C build all targets clean · phase4 lit 6/6 · full IR lit 356 passed / 0 failed / 55 unsupported (green again after the fixture fix) · focused integration across spectral + sharding + pipeline + MegaMoE + composition + dataflow 87 passed · full Python unit suite 12,754 passed / 0 failed (5,958 skip cleanly in a worktree without a build dir — the C++/device lanes were verified in the build tree above) · mypy 0 · ruff clean · dashboards regenerated and in sync.

Note on the incoming set's own log: it recorded that rg is unavailable on this host — /usr/bin/rg (ripgrep 14.1.0) is in fact installed.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 071a28265a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[gstoner]

All three findings verified and fixed in 5662c3e — the first one was worse than reported, and chasing it surfaced a fourth issue.

P1 (interleaved dependencies) — confirmed and quantified. On num_stages=4/num_chunks=2, 64 of 128 steps shared a key and 32 of 56 cross-stage forward steps lost their producer edge, so the carrier genuinely permitted a virtual stage to run before the stage feeding it. Now keyed by the virtual stage, with the stage count derived from the emitted schedule (num_stages × num_chunks — the old stage + 1 < num_stages bound was wrong for interleaving too). Zero missing producer edges across interleaved 2-chunk, 3-chunk, and both non-interleaved configs. The silent ordering filter that hid this now fails closed for the forward chain.

A deeper planner property this exposed, recorded rather than papered over: the interleaved generator emits backward steps in ascending stage order (num_stages=2/chunks=2, micro-batch 0: stage 0 B at clock 4 … stage 3 B at clock 7) — the opposite of gradient flow — so the backward-downstream edge is inexpressible in the emitted order. I omit it with a named comment instead of asserting an ordering the schedule doesn't realize; fixing the generator is its own change, not a review fix.

P1 (spectral fail-open) — confirmed. Fixed as you suggested: the producer now carries the consumed policy as tessera.spectral_semantic, and the Schedule→Tile consumer re-verifies workspace_bytes, native_entry, and normalization against it, failing closed on mismatch or absence. Pinned by a test that changes workspace_bytes while leaving both digests intact. Being precise about the limit: binding the payload into schedule_digest itself (so the pair can't be co-edited) requires extending the Schedule Object identity — recorded as follow-up, not claimed.

P2 (carrier validation) — confirmed and implemented. The legality pass now parses the rows and rejects duplicate ids, malformed rows, dangling dependencies, a producer ordered no earlier (cycle / producer-after-consumer), and stages or micro-batches disagreeing with the declared dimensions.

Fourth issue found while verifying: the spectral end-to-end tests were skipping everywhere — CI's unit lane is Python-only and a worktree has no build — so they had never actually executed. Run here against a real tessera-opt: 10/10, including the two new fail-closed proofs.

Gates: ninja all targets clean · phase4 lit 6/6 · full IR lit 356 passed / 0 failed · pipeline + sharding + spectral suites 58 + 10 with tessera-opt live · full Python unit suite 12,759 passed / 0 failed · mypy 0 · ruff clean.

🤖 Generated with Claude Code