release: 0.66.0 — the honest-GPU model release (PMAT-1096)

[noahgift]

Ticket

PMAT-1096 · milestone 0.66.0 · epic #2873

What this is

The 0.66.0 release bump, cut from the operator-curated milestone (2026-09-08/09): #3022 (F-1, #3050), #2971 (L0-1a/b, #3026 #3032, issue left open — see below) and #3063 (T0). The PP-066 obligation rows moved to 0.68.0 on 2026-09-09; no speed number ships in 0.66 (claims ratchet).

• scripts/bump-version.sh 0.66.0 — root workspace + the crates/facades/ upstream pins + facades lock (39 files); the facades' own version stays 0.4.0 by design (Publish provable-contracts{,-cli,-macros} as facade crates so existing users are carried forward, not stranded #2546)
• CHANGELOG.md gains [0.66.0] and the missing [0.65.1] / [0.65.2] sections
• docs/audits/impl-PMAT-1096-receipt.md (status: partial until the tag and the cascade), reduced gate logs, docs/audits/quorum-PMAT-1096-2971.json

Merge order (this PR merges LAST)

Draft until #3050 and #3063 are merged, so the tag carries them. Then ready + auto-merge — done by the release autopilot on lambda (/mnt/nvme-raid0/agent-wt/rel-066-autopilot/autopilot.sh, terminal states in STATUS beside it): wait for the two merges → gh pr ready + gh pr merge --squash --auto → dogfood --phase pre-publish on the merge commit (must be GO) → tag v0.66.0 + gh release create → check_publish_preflight.sh (must PASS) → cascade-drain.sh --target 0.66.0 → cargo install aprender --version 0.66.0 → dogfood --phase post-publish. Every irreversible step sits behind the repo's own gate; any refusal stops it with a STOP line.

Local gates at HEAD (receipt verification: rows, logs under docs/audits/impl-PMAT-1096-logs/)

gate	exit
cargo fmt --all -- --check	0
cargo deny check advisories	0
cargo test -p aprender-contracts --lib	0 (1501 passed)
make gate (pmat verify · guard_tree --no-cargo · cargo check --workspace --tests)	0
bash scripts/check_roadmap_diff_additive.sh	0 (added=1, the PMAT-1096 mint)

#2971 closed on a 3/3 quorum after the reporter's exact file was measured

Round 1 split 2 close / 1 keep-open (agreed=false, ask (3) and the Coder→base-Instruct step unmeasured). The reporter's exact GGUF (qwen2.5-1.5b-instruct-q4_k_m.gguf, sha256 6a1a2eb6…9407e) was then run on lambda with a cuda apr built from this tree: 78 positions, min cosine 0.9978 ≥ 0.98 (evidence/parity/l0-1/lambda/). Round 2: agreed=true, 3/3 close; closed through the T-6 gate with the citation (docs/audits/quorum-PMAT-1096-2971.json).

Pre-publish dogfood: NO-GO → GO on this branch

Run early on the branch head: five red rows, none from the bump — six shifted claim literals (shrink-only ratchet), a 7.1-day-old PP-26 witness, eight bashrs SEC/DET findings from #3032/BSE scripts, C14 run on a non-cuda binary (a tool defect), stray untracked dirs. Each fixed at the root (b891ab789); round 2 is GO (docs/audits/impl-PMAT-1096-logs/dogfood-pre-publish-b891ab789.json). The R5 receipt is re-taken on the merge commit by the autopilot.

After merge (orchestration, this session)

tag v0.66.0 → gh release create (notes = the CHANGELOG section) → scripts/dogfood.sh --phase pre-publish → scripts/check_publish_preflight.sh → scripts/cascade-drain.sh --target 0.66.0 → post-publish cargo install aprender --version 0.66.0 QA → receipt status: complete in a docs PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BjhtNUSensCYpQb3mCYLod

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3069 head=2794c18dfa83dd79207d9a4dd0391275caebee8d verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.