docs(adr): ADR-077 — pruning + distillation priority (Metal + CPU)

[oceanwaves630]

PR description authored by Claude (Anthropic agent) on behalf of @oceanwaves630.

ADR-077 — Pruning + Distillation Priority (Metal + CPU)

Eighth sequencing ADR in the bundle set (siblings ADR-070…076). Docs-only — one new file,
docs/adr/ADR-077-pruning-distill-priority.md. Follows the ADR-073/076 split-verdict house
standard: a Measured / source-verified table (rows tagged runtime-measured | unit-test/gate-pinned |
source-read, each with an origin/main pointer) held apart from an untrusted external prior-art
survey folded as data.

Split verdict — both families NEEDS-EXPERIMENT, at different maturity

• Pruning is shipped substrate, never validated. A Block-Influence scorer (pruning.rs, ADR-060
  P1, PR feat(inference): ShortGPT block influence scoring (ADR-060 P1) #133 merged), a real in-memory apply_layer_mask, and a Metal score_layer_importance that
  runs actual forward-pass activations and produces a type-balanced GDN/GQA mask are all
  shipped and wired. But the masks have never been PPL-gated — only a 4-prompt / 8-token
  logit-cosine smoke test — and no pruning surface is reachable from any binary. ADR-060's own status
  footer confirms the calibration pipeline, width pruning, and PPL-gated workflow are not shipped.

• Distillation is a false cognate. The one artifact named DistillationPipeline is a 6-class
  conversational-intent labeler with a hardcoded placeholder teacher call (issue tune: DistillationPipeline returns simulated labels, HTTP client not wired #11) — not knowledge
  distillation. There is zero KL/soft-target/teacher-logit machinery anywhere. Issue train(tune): KL-distillation recovery loop for pruned models on top of the existing LoRA NLL trainer #499's own
  body already states this ("DistillationPipeline is not relevant infrastructure... the reusable
  asset is the CPU LoRA NLL trainer"). That trainer is hardcoded to the 24-layer 0.8B stack
  (TOP_LAYER = 23) and cannot touch the 27B model without generalization.

Decision — validate the shipped scorer first; KD is gated behind it

• P1 — run the PPL-gate on the shipped scorer (cli(prune): expose score_layer_importance/LayerPruningPlan as a calibration+scoring command with a PPL gate #492). The pivot: wire score_layer_importance to
  a CLI + real calibration corpus + PPL gate + serialized PrunePlan. Reuses 100% of the shipped
  scorer and the existing eval_perplexity PPL math behind a thin --layer-mask flag. cli(prune): expose score_layer_importance/LayerPruningPlan as a calibration+scoring command with a PPL gate #492 blocks
  research(prune): validate block-influence ranking against post-pruning PPL on a hybrid GDN/GQA model #495/train(tune): KL-distillation recovery loop for pruned models on top of the existing LoRA NLL trainer #499/research(prune): validity of teacher-KV-prefix reuse for distillation recovery without re-prefill #502/research(prune): iterative re-scoring with post-prune KL feedback vs. one-shot block-influence #503 by its own text.
• P2 — validate BI-ranking vs PPL on the hybrid (research(prune): validate block-influence ranking against post-pruning PPL on a hybrid GDN/GQA model #495) — the ShortGPT-BI-on-hybrid question the
  survey flags as unresolved in the literature.
• P3 — generalize TOP_LAYER = 23 — a hard prerequisite for any 27B recovery/KD, currently
  untracked by any open issue (surfaced here to file).
• Gated, not ranked: train(tune): KL-distillation recovery loop for pruned models on top of the existing LoRA NLL trainer #499 (KL-recovery) needs a KL loss term (greenfield), the P3 refactor, and a
  pruned model worth recovering — gated behind P1/P2. research(prune): validity of teacher-KV-prefix reuse for distillation recovery without re-prefill #502/research(prune): iterative re-scoring with post-prune KL feedback vs. one-shot block-influence #503 conditional/secondary.
• Conflation guard: DistillationPipeline/tune: DistillationPipeline returns simulated labels, HTTP client not wired #11 is intent classification, not compression KD; the
  standing teacher-proxy distillation project is greenfield here.

This ratifies the experiment-gated discipline of ADR-073/076, with one twist: the experiment here
validates an already-shipped primitive (the scorer) rather than deciding whether to build one.
Builds on ADR-060 (honest partial-shipment) and amends nothing — the distillation false-cognate is
already corrected in #499's body and audited by closed issue #302.

Verification

Every load-bearing pointer re-verified against origin/main @ 58e885ada: BI scorer
(pruning.rs:41,59, lib.rs:76, PR #133 merged), apply_layer_mask:638,
score_layer_importance:12397 + type-quota :12437-12456, DistillationPipeline placeholder
(distill.rs:69), zero KL machinery (grep, zero hits), TOP_LAYER=23 (train_core.rs:13), ADR-060
status footer (:680-684), and all issue states (#492/#495/#499/#502/#503/#11 OPEN, #133 MERGED,
#445/#302 CLOSED). The external survey is genuinely repo-grounded (cites a verified ancestor commit)
and corroborates every finding.

Docs-only; no code paths change.

[ohdearquant]

SPEC-GATE: SIGNED (Leo). All three press-time gates verified at content level before pressing: the false-cognate row pins to distill.rs:69's placeholder-teacher line and IntentLabels with #499's own body as corroboration, #730 is filed and OPEN with the correct name (verified directly) and referenced at R10/P3/consequences, and the type-balanced-scorer claim cites the quota loop at metal_qwen35.rs:12437-12456. The validates-shipped-primitive framing is a genuine addition to the bundle's epistemic vocabulary — an unvalidated shipped scorer is neither greenfield nor proven, and P1 (#492 wiring + PPL gate) is correctly the single unblocking measurement. Surfacing-then-FILING the TOP_LAYER=23 blocker is the ADR finishing its job. Bundle H signed — the sequencing set D/B/A/E/F/G/H is complete. Proceed to I (multimodal deferral ADR), then the P2 dispatch-count-cut build lane.