docs(adr): ADR-076 — adaptive reasoning-budget priority (Metal + CPU)

[oceanwaves630]

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

ADR-076 — Adaptive Reasoning-Budget Priority (Metal + CPU)

Seventh sequencing ADR in the bundle set (siblings ADR-070…075). Docs-only — one new file,
docs/adr/ADR-076-adaptive-reasoning-priority.md. Follows the ADR-073 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.

The verdict is a split at two maturity levels

• Static s1 reasoning-budget forcing is shipped, not greenfield. GenerateConfig.reasoning_budget
  force-injects </think> after N reasoning tokens, wired on both the CPU decode loop and the
  Metal candidate-sampling path, grammar-fail-closed and mutation-verified (PR test(inference): grammar × reasoning-budget combined-path fail-closed test #511), zero cost
  when disabled (PR feat: Studio macOS redesign + s1 reasoning-budget engine support #435 A/B). Its only open sub-lever is surface coverage — it is reachable on
  chat_metal and lattice_serve but not on the public lattice CLI. That is standard plumbing,
  rankable today with zero new data.

• Adaptive / entropy-gated budgeting is measurement-gated. Zero lattice code, zero committed eval
  data, no margin/entropy instrumentation. It is an already-filed, unstarted 6-issue family
  (research(serve): confidence-gated adaptive reasoning + local safety triage (tracking) #482 tracking → eval(serve): does a scratchpad improve label-token calibration (ECE) at 0.8B? #487/eval(serve): token-margin/entropy AUROC vs. reasoning correctness at 0.8B (general-domain) #493/eval(inference): MTP draft-head disagreement as a self-consistency / uncertainty proxy #496/eval(gdn-state): correlate state-observer signals against reasoning correctness (fallback uncertainty oracle) #497 → feat(serve): entropy-gated adaptive reasoning budget — dynamic s1 (extends #435) #500). eval(serve): token-margin/entropy AUROC vs. reasoning correctness at 0.8B (general-domain) #493 (token-margin/entropy AUROC vs. correctness,
  G1 offline, promote ≥0.75 / kill <0.65) is the defined minimal experiment and blocks feat(serve): entropy-gated adaptive reasoning budget — dynamic s1 (extends #435) #500 by its
  own text.

Decision — rank the experiment, not the unbuilt feature

• P1 — run eval(serve): token-margin/entropy AUROC vs. reasoning correctness at 0.8B (general-domain) #493 (the AUROC measurement; the pivot that unblocks the family).
• P2 — run eval(serve): does a scratchpad improve label-token calibration (ECE) at 0.8B? #487 (scratchpad→ECE calibration; parallel G1 experiment).
• P3 — expose the shipped static budget on the public lattice surface (certain-value low-risk
  plumbing; no tracking issue filed yet).
• Gated, not ranked: feat(serve): entropy-gated adaptive reasoning budget — dynamic s1 (extends #435) #500 graduates to a build priority only if eval(serve): token-margin/entropy AUROC vs. reasoning correctness at 0.8B (general-domain) #493 clears AUROC ≥0.75;
  honest kill clause — if it can't beat the best fixed static budget, the shipped static budget stays
  the default. eval(inference): MTP draft-head disagreement as a self-consistency / uncertainty proxy #496/eval(gdn-state): correlate state-observer signals against reasoning correctness (fallback uncertainty oracle) #497 are conditional/secondary on eval(serve): token-margin/entropy AUROC vs. reasoning correctness at 0.8B (general-domain) #493.

This applies the experiment-gated-priority discipline from ADR-073: a lever family with zero
local-measured data is gated behind a defined minimal experiment rather than ranked on literature
confidence — here the experiment is already filed as #493 with its own thresholds, so the ADR
ratifies that gate. No prior ADR is amended — this is the first ADR in the reasoning area.

Verification

Every load-bearing pointer was re-verified against origin/main @ 4cb006d32 (the recon and this
author both read a throwaway worktree at that head, no stale-main risk): the shipped field + decision
functions + full test set (qwen35_config.rs:690,747-780,1635-1743), both-backend wiring
(generation.rs, metal_qwen35.rs:13131-13148), PR #511 mutation-tested fail-closed, enable_thinking
hardcoded true in serving binaries with zero hits in the public CLI, the capability-matrix gate
(make lint-docs), and all seven issues confirmed OPEN. PR #435 (MERGED 2026-06-29) + #511 (MERGED
2026-07-01).

Docs-only; no code paths change.

[ohdearquant]

SPEC-GATE: SIGNED (Leo). Read the split boundary at content level before pressing: ranking the filed experiment (#493, with its own promote/kill thresholds) above the feature it gates is the ADR-073 discipline correctly ratified, the kill clause names static-stays-default in so many words, #496/#497 are properly conditional on #493's outcome, and the logprobs/best-of-N conflation fence is a good inclusion. Declining to file P3 until it ranks and declining to invent an amendment where nothing needs correcting are both the right restraint. Bundle G signed — D/B/A/E/F/G done, H and I remaining, then the P2 dispatch-count BUILD lane.