docs(adr): ADR-079 adaptive micro-LoRA brain (composed train-govern-compose-route-consume loop)

[oceanwaves630]

Summary

ADR-079 is the synthesizing record for the composed multi-adapter LoRA system built one stage at a time across six independently-merged PRs. No prior document maps the loop as one system: seven existing LoRA ADRs each cover a single-adapter lifecycle slice or an orthogonal composition axis, and ADR-056's own appendix admits its designed training pipeline was never built.

The ADR names shipped / partial / gap per stage of the TRAIN → GOVERN → COMPOSE → ROUTE → CONSUME loop, every load-bearing row anchored to a merge commit + origin/main file:line.

Composing PRs

• feat(tune): exact-gradient LoRA backward + multi-layer training pipeline #191 exact-gradient CPU backward trainer (TRAIN)
• feat(tune): governed LoRA manifest + fail-closed loader + micro-LoRA trainer #444 governed manifest + fail-closed loader (GOVERN)
• feat(inference): weighted LoRA adapter mixture at decode time (zero kernel change) #443 decode-time weighted mixture blend, zero kernel change (COMPOSE)
• feat(fann): online router refit (single-sample policy gradient + EWC++) with convergence bench #448 + feat(tune): add online router-refit from preference feedback (update_router) #453 RLOO + Fisher online router-refit (ROUTE)
• feat(inference): add query-likelihood reranking to the Qwen3.5 decoder #718 causal-LM query-likelihood rerank; ADR-057-D1 cross-encoder rerank (CONSUME)

Two record corrections

1. The merged CPU backward trainer is feat(tune): exact-gradient LoRA backward + multi-layer training pipeline #191 (7b46f6e77), not commit f34ed3b1. The "surface-B" GDN LoRA weight-grad commits are not ancestors of origin/main: f34ed3b1 was in the now-closed PR feat: LoRA backward training (GQA+GDN) + LatticeStudio macOS instrument app #193, and the current standalone PR feat(tune): surface-B GDN LoRA weight gradients + train_grad_full fields #202 (different head commits) is still OPEN — neither is on head. So today only GQA q_proj/v_proj adapters train; GDN LoRA weight-gradients do not exist in mergeable code.
2. feat(inference): add query-likelihood reranking to the Qwen3.5 decoder #718 is unrelated to Stable rerank API on lattice-inference with LoraHook injection #59. feat(inference): add query-likelihood reranking to the Qwen3.5 decoder #718 is the causal-LM reranker; Stable rerank API on lattice-inference with LoraHook injection #59 (cross-encoder LoraHook injection) was addressed by PR feat: implement ADR-057 LoRA full-lifecycle consumer API (D1-D5) #65 five weeks earlier. Two independent rerank-via-seam consumers now exist; the ADR names them as siblings.

Named gaps (each with a re-entry trigger)

G1 GDN LoRA weight-grads (#202), G2 in-process route→refit→route loop-closure, G3 EWC anchor-pullback as a live anti-forgetting mode, G4 rerank-consumer disambiguation.

Scope / review

• Docs-only. No code, no bench, no parity impact.
• Every reality row is source-read or PR/issue-state against origin/main @ b0604722d; the file:line anchors and the two corrections are the load-bearing claims to audit.
• INDEX note: this PR adds the ADR-079 row. ADR-071..078 are absent from INDEX.md (pre-existing — the bundle-sequencing ADRs were never registered); left for a separate housekeeping pass rather than scope-creeping this docs PR.

[ohdearquant]

SPEC GATE SIGNED: ADR-079. Composed-loop map with file:line pins at b060472, two load-bearing record corrections (GDN grads not on main — PR #202 open; #718 does not close #59 — that was #65), four named gaps each with re-entry triggers, S-row rider. Canonical map for Lane 1; supersedes nothing.