docs: bind AUTODIFF_NEXTGEN_PLAN into the compiler queue and layout record

[gstoner]

Follow-up to #582 (merged). That PR added the design document; this one applies the bindings it proposed, so the queue and the layout record actually point at it. Three files, docs only — no code, no generated docs, no status claims.

INTEGRATED_COMPILER_PLAN.md

• W6.3 — replaces research estimate required with the design summary and phased estimates, naming AUTODIFF_NEXTGEN_PLAN.md as the owning design doc: AD-LAW-1 (~1–1.5w, host-free, no dependencies — may run in parallel like Orders 12–13), AD-WEIL-1 (~3w, host-free), AD-JET-STRUCT-1 (~4w), AD-JET-IR-1 (~5–6w, open — depends on W4-PRODUCT-1, LAYOUT-ALG-1 L1/L5, real batching_rules, and the S5 numeric_policy carrier), AD-OPERATOR-1 (~2w, independent — absorbs S8's unowned implicit-diff strict-complementarity item), AD-CERT-1 (consumer-gated). Estimate: ~10w excluding AD-JET-IR-1.
• W6 exit — gains the plan's falsifiable stop condition: if the k+1-vs-2ᵏ curve does not separate by k=3, the IR descent is not worth its cost and the program stops at the reference + law infrastructure.
• W6.4 note — the shared-substrate reuse hypothesis stays a hypothesis, now as an explicit AD-WEIL-1 acceptance criterion (Cl(3,0) over the same multiplication-table substrate, ga/signature.py as oracle) instead of a sequencing assumption.
• E2E-REAL-6 — records the Law-3 adjoint sweep as a pending gate, explicitly not active until AD-LAW-1 exists, and states that it completes the existing pointwise numeric identity on the transpose axis rather than replacing per-family derivative oracles (a matched-wrong JVP/VJP pair passes the adjoint identity — see the plan's §3.5 caveat).

CUTE_IR_ASSESSMENT.md

• Adds the jet coefficient axis as the sixth layout-algebra consumer in §3: attach = logical_product, planar-vs-interleaved = mode regrouping, footprint = cosize, fusion legality = the same ⊑ factorization FORGE asks for, digest equality = coalesce, emitted index math = crd2idx. Written against today's substrate (two-valued LayoutOrder enum + ~25 string-template index sites) a jet emitter would hand-roll coefficient indexing per backend — the duplication L4 exists to consolidate.
• Updates the three prose counts that the new row changes (§0 bottom line, §3 lead-in, §4 "For"), and notes this row is different in kind: it was found during design, before any code existed — the argument working prospectively rather than as archaeology.

AUTODIFF_NEXTGEN_PLAN.md

Binding-status bookkeeping only: §5 now marks bindings 1 and 5 accepted (2–4 still proposed), §6 records the CUTE row as applied, and the routing banner states that no slice has started — every status word in the plan is design intent, not implementation.

Verification

tests/unit/test_audit_docs.py 11 passed; scripts/check_generated_docs.sh ok (25 in sync). Host: Mac (host-independent doc-only change).

Also in this cleanup

Deleted five fully-merged branches (local agent/autodiff-nextgen-plan, worktree-apple-lowp-norm-vjp; remote agent/autodiff-nextgen-plan, agent/central-compiler-tsol-tools-plan, agent/e2e-real-6-stateful-authority, agent/w4-product-1-multiblock-regions) after confirming each was merged into main and no worktree was checked out on them.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 61299757af

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".