Spec amendment 1: typed RIR semantics + D21 (no string-tagged dispatch)

[StarGazerM]

Summary

First spec refinement after PR #74 (the redesign baseline). Addresses 8 semantic gaps in the RIR vocabulary where operational meaning was hidden behind string-tagged fields instead of typed ops. Expands RIR from ~10 op families to 30 typed ops + 5 carrier dataclasses.

What's new

§ 1 ACID sub-clause

Every operational semantic — termination, cross-stratum deps, delta-variant ordering, multi-head emission, kernel fusion, pragma compatibility — must be expressed as typed IR ops or typed metadata, not as string-tagged dispatch in renderers. Renderers may dispatch on op TYPE, never on op string fields.

§ 3.1.1.1 expanded RIR vocabulary (30 typed ops, 5 carriers)

Group	Count	Notable ops
Structure	4	RunnerStruct, StepDispatch, FixpointLoop, Step (with explicit deps: tuple[StepId, ...] DAG edges)
Termination (new typed union)	4	ConvergenceCheck, MaxIterationFuel, EitherFirst, AndAll
Scheduling	5	KernelLaunch + 4 schedulers
Kernel bodies (was string-tagged phase)	3	CountKernelDef, MaterializeKernelDef, FusedKernelDef
Variant ordering	1	DeltaVariantSet(ordering=Sequential|Independent|Fusable)
Multi-head	1	MultiHeadInsert(heads, ordering=Sequential|Concurrent)
Dedup	1	DedupTable
Maintenance (was 1 string-lump)	11	distinct typed ops: ComputeDeltaCall, MergeIndexCall, RebuildIndexCall, CheckSizeCall, etc.

Plus carriers: HeadSpec, ViewBinding, ComposabilityMeta, VariantOrdering, HeadOrdering.

§ 7.1 D21 discipline rule

No @register_render body may dispatch on a string field of its op. AST scan of render modules; fail on if op.kind == ... / match op.kind: patterns.

§ 10 Amendment log

Acknowledges the spec is a living document during the redesign; subsequent amendments are PR-shaped and reviewed.

Two corrections surfaced during review

1. Maintenance op count was wrong — the original spec said "14 maintenance ops" but actual count from orchestrator.py:506-588 is 11 (verified against mir/types.py). The spec conflated ExecutePipeline + ParallelGroup + a duplicate branch into the count. Now corrected throughout.
2. KernelDef(name, phase: KernelPhase, body) is superseded — replaced by 3 typed ops. Old code block left intact with a supersession callout, per the "don't rename/move existing headers; expand instead" rule.

PR-2 split: YES (6 → 7 PRs total)

With 30 typed ops + 11 lowerings to maintenance ops, PR-2's scope ballooned past comfortable. Split:

• PR-2a = framework + structural (4) + scheduling (5) + kernel-body (3) + dedup (1) = ~2400 LOC
• PR-2b = maintenance (11) + termination (4) + variant ordering (1) + multi-head (1) + view bindings = ~2100 LOC

Both gate on verify_runner_completeness (WARN in 2a, HARD in 2b).

Final ledger: PR-1 → PR-2a → PR-2b → PR-3 → PR-4 → PR-5 → PR-6. Strictly serial.

Test plan

• sphinx -W build succeeded
• ruff check + format (CI v0.9.10): clean
• No markdown links to local .md files in doc body (only the existing example of the forbidden pattern in a code span at line 14)

🤖 Generated with Claude Code