docs/plans: execution-spine design — primitives → DependencyView → realization (draft for sign-off)

[gunbai-bot]

Design draft (docs-only, +142 lines, no code) for the execution spine: run ≜ realize ∘ materialize ∘ dependency_view — making DependencyView a total function over the primitives and realization/materialization its only downstream readers, so "runs as fast as the hardware allows" and "a model connected to nothing is unwritable" hold by construction.

Why now

This is the operational follow-on to the §0 roadmap spine (#6343): the fracture it fixes is verified live on main (re-verified today at 24c2be4):

• Two scheduling authorities coexist — std.realization family (CI floor) and std.dependency/topological_layers (executor, runtime_run, self-host emit); neither derives from the other, and the floor imports both and hand-builds a RealizationPlan (ci_floor_plan.dag:441).
• The shard subsystem is wired to the wrong purpose — proven-total partition, consumed only for selection/memory-sizing, never parallel execution.
• The compiler compiles at 1 of 128 cores — serial catamorphism, no parallelism in the infer path (grep-verified), >1h whole-tree serial where frontend+normalize is 3s.

What the draft claims

• §1–§2: dependency_view as THE canonical total reading; realize/materialize as its only readers; single run authority with bypasses deleted (§4).
• §3/§3′: parallelism as a theorem of the structure — catamorphism (fractal) + monoidal combine (parallelizable) + edge-completeness (sound); the 1/128 is non-monoidal accumulator-threading, not a limit.
• §5: sharding subsumed, not deleted (totality proof carried forward).
• §6: two realization scenarios, single-process-multi-thread first; resource-aware packing on measured peaks (first live instance: the floor's own gate-series, 4.8 GiB of 33 GiB budget → SUM(gates) collapses to compile-clean + MAX(gate)).
• §7: the in-flight lanes (namespace-only resolution, SymbolIndex, Lever a slice 2: reground selection on DependencyView #6335 DependencyView, floor scheduler) already are the spine's parts — this names what they converge on.
• §9: measurable acceptance — core-utilization receipt vs critical-path floor, discriminating serial-chain RED, spine-coverage RED, byte-identical materialization oracle.

Sign-off requested (§10)

FLAG A (run composition + single authority) · FLAG B (realization grain: per-declaration) · FLAG C (shard subsume-not-delete) · FLAG D (std.dependency fundamental, std.realization repositioned as reader) · FLAG E (single-process-multi-thread first).

Test plan

• Docs-only; no code or CI surface touched. Evidence claims in §0 re-verified against main @ 24c2be4 today (floor dual-import + hand-built plan at ci_floor_plan.dag:441-450; no rayon/par_iter/spawn in v1_compiler_infer.rs).

🤖 Generated with Claude Code