release: 0.71.0

[noahgift]

Release bump for 0.71.0: bump-version.sh 0.71.0 across every workspace (--check green), the CHANGELOG section below, and the CRUX prompt certification carried to evidence/crux/0.71.0/ from 0.70.3 (apr_commit set to this bump's parent).

Cut by hand at its parent, not by prepare_bump.sh: that script cuts only at a commit the T-2 preflight judged GO, and T-2 has never produced a GO, so it is not a gate for this train. This PR is opened unarmed.

The 0.71.0 train: release checks move into code with a fixed stage budget, checks that cannot block a merge or a release run nightly instead of on the PR path, a nightly release rehearsal and nightly train judge one commit end to end, the CI helpers and the CRUX judge are ported to Rust, the Laya ModernBERT encoder lands, and Qwen3.5 hybrid GGUF export loads in llama.cpp. From this release on, a release ships on CRUX smoke; larger model rows run nightly and each red row is a ticket.

Added

• feat(release): nightly assets rehearsal — build every release asset on a schedule, unable to upload (0702-P3a: nightly assets rehearsal for RR-P03/P04 — build every release asset at main's head on a schedule, never upload, no secrets (child of #4678) #4806) (feat(release): nightly assets rehearsal — build every release asset on a schedule, unable to upload (#4806) #4813)
• feat(ci-tools): aprender-ci-tools crate — Rust ports of the CI helpers (PY-PORT-2, split 1/2 of PY-PORT-2 (PMAT-4741): port the CI Python tools to aprender-ci-tools (C301) #4847) (feat(ci-tools): aprender-ci-tools crate — Rust ports of the CI helpers (PY-PORT-2, split 1/2 of #4847) #4916)
• feat(crux-judge): Rust CRUX judge (PMAT-4802, split 2/2 of PY-PORT-2 (PMAT-4741): port the CI Python tools to aprender-ci-tools (C301) #4847) (feat(crux-judge): Rust CRUX judge (PMAT-4802, split 2/2 of #4847) #4917)
• feat(hooks): pre-push tag guard, accepting a release tag at exactly its release branch head (feat(hooks): pre-push tag guard, accepting a release tag at exactly its release branch head #4944)
• feat(release): the stage list lives in code - one data file, one reader, a budget self-test (merge 9, tag 5, publish 5) (feat(release): the stage list lives in code - one data file, one reader, a budget self-test (merge 9, tag 5, publish 5) #4967)
• feat(release): nightly release rehearsal on one commit, with the three-night streak (feat(release): nightly release rehearsal on one commit, with the three-night streak (C332 a) #4997)
• feat(export): Qwen3.5 hybrid GGUF export that llama.cpp loads (feat(export): Qwen3.5 hybrid GGUF export that llama.cpp loads #4998)
• feat(code): apr code's JSON document reports the device it ran on (Refs apr code on Qwen3.5 Q4_K_M (4B, 9B) on CUDA: does it complete a scripted edit-and-verify task on lambda and gx10? UNMEASURED, and its driver runs through apr serve --gpu (#3571) #3719) (feat(code): apr code's JSON document reports the device it ran on (Refs #3719) #5005)
• feat(contracts): findings-ledger-v1, a findings line under contract (APR-EPIC-001 A8: intake control in-tree — findings-ledger-v1, CLAUDE.md rewrite, check_pr_closes.sh #4455) (feat(contracts): findings-ledger-v1, a findings line under contract (#4455) #5006)
• feat(serve): X-Trace-Level step|layer runs the tracer, and says measured only when it ran (APR-OBS-001 OBS-09: serve tracing: X-Trace-Level runs inference_trace (first row of 0.71 serve epic) #4496) (feat(serve): X-Trace-Level step|layer runs the tracer, and says measured only when it ran (#4496) #5015)
• feat(ttft): the V1 TTFT ratio verdict, the TTFT split by phase, and a baseline script that refuses an untrustworthy run (Refs V1 (0.71): TTFT ratio vs llama.cpp with a 95% lower bound — baseline on both GPU hosts #4954) (feat(ttft): the V1 TTFT ratio verdict, the TTFT split by phase, and a baseline script that refuses an untrustworthy run (Refs #4954) #5037)
• feat(dogfood): write a dogfood-receipt/v1 and run the gate path with no Python (Adopt dogfood-receipt-v1 in apr-dogfood, and remove python3 from it (ARB-SELF-001 ASL-05) #4377) (feat(dogfood): write a dogfood-receipt/v1 and run the gate path with no Python (#4377) #5065)
• feat(release): the pass's runner-wait rule reads and prints inside its own waits (E1 D6: release pass runner rule — 10-min hold, 60-min cap, read+print only (C354 3, C357 1) #4929) (feat(release): the pass's runner-wait rule reads and prints inside its own waits (#4929) #5078)
• feat(nightly-train): a red or not_measured night opens or updates one ticket with one owner (nightly train: a red night opens a ticket with an owner within 60 min #5081) (feat(nightly-train): a red or not_measured night opens or updates one ticket with one owner (#5081) #5085)
• feat(decide): carry feat(decide): the Laya ModernBERT encoder and decision library (first slice of #4634) #4941, the Laya ModernBERT encoder and decision library (feat(decide): carry #4941, the Laya ModernBERT encoder and decision library #5101)
• feat(nightly): dogfood and readiness lanes get their producers on main (feat(nightly): dogfood and readiness lanes get their producers on main #5103)
• feat(bench): V2 serve receipt — load_ms + TTFT beside pp512/tg128 (Refs V2: apr serve keeps the model loaded; Rust receipt reports load_ms and TTFT beside pp512/tg128 (APR-071) #4953) (feat(bench): V2 serve receipt — load_ms + TTFT beside pp512/tg128 (Refs #4953) #5106)

Fixed

• fix(release): deep-nodefault no longer reads green on a failing cargo exit (fix(release): deep-nodefault no longer reads green on a failing cargo exit #4874)
• fix(release): T-2 can return GO on main's head — defer only the pre-bump CRUX ladder row (D5) (fix(release): T-2 can return GO on main's head — defer only the pre-bump CRUX ladder row (D5) #4931)
• fix(release): the coverage version-only rule admits the carried CRUX prompt certification (P7 D3) (fix(release): the coverage version-only rule admits the carried CRUX prompt certification (P7 D3) #4932)
• fix(release): T-1 dogfood ladder gate waits for the models lane and reads its CRUX receipts (release: under the CRUX policy the T-1 dogfood ladder gate runs before the models lane has written its receipts #4930) (fix(release): T-1 dogfood ladder gate waits for the models lane and reads its CRUX receipts (#4930) #4933)
• fix(release): the models lane's remote leg fetched only main, so a patch release could not build there (fix(release): the models lane's remote leg fetched only main, so a patch release could not build there #4938)
• fix(gpu): a shared cubin temp file let a reader load a partial cubin (FLAKE: intermittent SIGSEGV in libcuda during gdn device tests (aprender-gpu --lib) #4956) (fix(gpu): a shared cubin temp file let a reader load a partial cubin (#4956) #4964)
• fix(gpu): cuBLAS tests synchronized the context inside another test's capture (FLAKE: intermittent SIGSEGV in libcuda during gdn device tests (aprender-gpu --lib) #4956) (fix(gpu): cuBLAS tests synchronized the context inside another test's capture (#4956) #4966)
• fix(gpu): a test filling the device let neighbours' allocations OOM in-process (FLAKE: intermittent SIGSEGV in libcuda during gdn device tests (aprender-gpu --lib) #4956) (fix(gpu): a test filling the device let neighbours' allocations OOM in-process (#4956) #4970)
• fix(guards): pin origin/main to the run's build base, so a re-run judges the main it was built on (fix(guards): pin origin/main to the run's build base, so a re-run judges the main it was built on #4996)
• fix(dogfood): the release model gates read a declared model list, not the model dir (fix(dogfood): the release model gates read a declared model list, not the model dir #5000)
• fix(test): MUT-05/06/07 assert the nightly mutants job; m001/m008 nested build drops CONTRACT_* env (fix(test): MUT-05/06/07 assert the nightly mutants job; m001/m008 nested build drops CONTRACT_* env #5004)
• fix(ci): libm-ban-live prints clippy's output when clippy fails for another reason (fix(ci): libm-ban-live prints clippy's output when clippy fails for another reason #5020)
• fix(ci): fat_driver case-table fixture lines never reach the runner as annotations (Refs fat_driver self-test leaks fixture ##[error] lines as real job annotations (read as a 12000s runner timeout) #4596, section waiter: self-test fixture prints '##[error] not completed within 12000s' 30 s into a job, masking the real failure (L19) #4610) (fix(ci): fat_driver case-table fixture lines never reach the runner as annotations (Refs #4596, #4610) #5029)
• fix(flake): disable test_stack_unwinding_does_not_crash under FLAKE-0 (fix(flake): disable test_stack_unwinding_does_not_crash under FLAKE-0 #5035)
• fix(guards): crux serve guard's EXIT trap runs only in its own shell (GATE flake: check_crux_serve_code.sh T3/T4 fail in the merge queue (run 38010593439) #5012) (fix(guards): crux serve guard's EXIT trap runs only in its own shell (#5012) #5036)
• fix(qwen35 CUDA-vs-CPU e2e logits exceed LOGITS_BUDGET 7e-2 on GB10 sm_121 (7.105e-2, 2/2 deterministic; budget measured on 4090 only) #4523): per-arch qwen35 LOGITS_BUDGET + a nightly sm_121 lane for the parity module (fix(#4523): per-arch qwen35 LOGITS_BUDGET + a nightly sm_121 lane for the parity module #5038)
• fix(tests): four cbtop tests passed while apr-cli did not compile (fix(tests): four cbtop tests passed while apr-cli did not compile #5042)
• fix(cuda): drop graph-mode buffers after a failed decode-graph capture (C14) (fix(cuda): drop graph-mode buffers after a failed decode-graph capture (C14) #5043)
• fix(release): the fleet cells gate judges in bash and awk, not python (ARB-AUD-10: fleet_cells_gate.sh lands on main without Python, after ARB-AUD-01 #4350) (fix(release): the fleet cells gate judges in bash and awk, not python (#4350) #5049)
• fix(nightly): ci-pending is asked again the same night, then ends RED (Nightly: ci-pending skips every build and ends green; retry the same night, else red #5053) (fix(nightly): ci-pending is asked again the same night, then ends RED (#5053) #5054)
• fix(ci): section git fetches read up to 3 times through fetch_p6.sh (P6) (ci: fence fat_driver case-table annotations; section git fetch reads up to 3 times (P6) #4936) (fix(ci): section git fetches read up to 3 times through fetch_p6.sh (P6) (#4936) #5066)
• fix(ont): rebind or delete the ghost binding rows; allowlist 163 -> 3 (ONT: rebind 229 ghost bindings in car/0.70.0's 9 new binding registries (shrink-only baseline) #4470) (fix(ont): rebind or delete the ghost binding rows; allowlist 163 -> 3 (#4470) #5067)
• fix(examples): examples-nightly 36 rows to pass or a skip class, fail 0 timeout 0 (examples-nightly: 29 examples fail and 7 time out on main (run 37169058144) #4756) (fix(examples): examples-nightly 36 rows to pass or a skip class, fail 0 timeout 0 (#4756) #5068)
• fix(guards): crux serve guard's stub server holds no caller fd and never outlives its case (check_crux_serve_code.sh: mutant M10 (m_t1) leaves its TERM-ignoring stubborn() server alive after every run; the stub inherits the caller's flock fd and deadlocks the lock queue #5019) (fix(guards): crux serve guard's stub server holds no caller fd and never outlives its case (#5019) #5073)
• fix(ont): binding rows after the ghost rebind — dup row, stale signatures, partial fits (ONT: rebind 229 ghost bindings in car/0.70.0's 9 new binding registries (shrink-only baseline) #4470 follow-up) (fix(ont): binding rows after the ghost rebind — dup row, stale signatures, partial fits (#4470 follow-up) #5077)
• fix(train): judge the commit of the night's scheduled nightly.yml run, not main's head at read time (fix(train): judge the commit of the night's scheduled nightly.yml run, not main's head at read time #5086)
• fix(dogfood): pinning guard scopes a gate path whole; receipt test covers the exit (Adopt dogfood-receipt-v1 in apr-dogfood, and remove python3 from it (ARB-SELF-001 ASL-05) #4377 follow-up) (fix(dogfood): pinning guard scopes a gate path whole; receipt test covers the exit (#4377 follow-up) #5090)
• fix(release): gates and train judge the producer's commit, not main's head (yokoten: every gate that judges main's head at read time #5082 hits 2,3,4,5,8,12) (fix(release): gates and train judge the producer's commit, not main's head (#5082 hits 2,3,4,5,8,12) #5094)
• fix(package): three crates gitignored a source file they ship (fix(package): three crates gitignored a source file they ship #5095)
• fix(models-nightly): no_token tests the credentials readable, not present (Release rehearsal: pick finds no nightly-evidence branch: the models crux producer never ran #5062) (fix(models-nightly): no_token tests the credentials readable, not present (#5062) #5097)
• fix(nightly): every registered lane names a producer on the commit, as its own check (nightly train: self-test refuses a row with no producer on main #5080) (fix(nightly): every registered lane names a producer on the commit, as its own check (#5080) #5104)
• fix(scripts): kill only the pid recorded at spawn, never a name pattern (fix(scripts): kill only the pid recorded at spawn, never a name pattern #5112)
• fix(run): qwen35 --gpu tok/s no longer bills the F2 guard as generation (qwen35 --gpu tok/s counts the CUDA build + F2 guard as generation (inference_ms not generation-only) — sibling of #3714 #4342) (fix(run): qwen35 --gpu tok/s no longer bills the F2 guard as generation (#4342) #5115)
• fix(nightly): the assets lane names its producer, the nightly assets rehearsal (fix(nightly): the assets lane names its producer, the nightly assets rehearsal #5116)
• fix(rehearse): pick walks 200 first-parent commits, and the verdict prints C's age (Release rehearsal: pick finds no nightly-evidence branch: the models crux producer never ran #5062) (fix(rehearse): pick walks 200 first-parent commits, and the verdict prints C's age (#5062) #5129)
• fix(silicon): an API read failure is NOT_MEASURED, and gh's stderr is kept (check_silicon_coverage.sh: an API read failure must be not_measured, and stderr must be kept #5031) (fix(silicon): an API read failure is NOT_MEASURED, and gh's stderr is kept (#5031) #5130)
• fix(cuda): Qwen3.5 batched prefill f32 dequants for IQ2_S, IQ2_XXS, IQ3_S, IQ3_XXS and Q2_K (fix(cuda): Qwen3.5 batched prefill f32 dequants for IQ2_S, IQ2_XXS, IQ3_S, IQ3_XXS and Q2_K #5131)
• fix(train): a stale pin is RED, and the lanes run counts for the commit it asserted (Refs B1 C1: nightly train judges the scheduled nightly.yml run's commit, not main's head at read time #5083, nightly_train: match scheduled release-lanes runs by Assert-HEAD C, not head_sha (#5082 follow-up) #5087) (fix(train): a stale pin is RED, and the lanes run counts for the commit it asserted (Refs #5083, #5087) #5132)
• fix(train): the self-test runs an unpinned copy, so it is green inside an installed bundle (Refs B1 C1: nightly train judges the scheduled nightly.yml run's commit, not main's head at read time #5083) (fix(train): the self-test runs an unpinned copy, so it is green inside an installed bundle (Refs #5083) #5136)
• fix(decide): inherit aprender-core from the workspace; guard that path pins name their target version (assets rehearsal: aprender-decide pins aprender-core 0.70.2 under a 0.70.3 workspace, so the rc stamp cannot resolve it #5121) (fix(decide): inherit aprender-core from the workspace; guard that path pins name their target version (#5121) #5137)
• fix(andon): read the day dir, not cache/: a printed verdict was reported as missing (fix(andon): read the day dir, not cache/: a printed verdict was reported as missing #5141)
• fix(night): a dispatched run on main counts as the night, one rule at every site (fix(night): a dispatched run on main counts as the night, one rule at every site #5142)
• fix(rehearse): wiring counts the train's distinct capture names, so the lanes read no longer fails pick (Refs Release rehearsal: pick finds no nightly-evidence branch: the models crux producer never ran #5062) (fix(rehearse): wiring counts the train's distinct capture names, so the lanes read no longer fails pick (Refs #5062) #5143)
• fix(serve): Qwen3.5 host-RAM guard + 1.27x host footprint, recalibrated (Refs Small unified-memory CUDA devices: 16 GiB headroom refuses every model, no refusal before OOM, wrong pull hint #4947) (fix(serve): Qwen3.5 host-RAM guard + 1.27x host footprint, recalibrated (Refs #4947) #5146)
• fix(decide): pack_laya runs a deterministic smoke path with no arguments (fix(decide): pack_laya runs a deterministic smoke path with no arguments #5157)
• fix(release): models gates on crux; ladder and six night rows go to LAB (fix(release): models gates on crux; ladder and six night rows go to LAB #5158)

Changed

• release: policy reader refuses an unquoted #; nightly rows match inside the entry (release: policy reader refuses an unquoted #; nightly rows match inside the entry #4923)
• release: 0.70.3 — streaming tool calls (release: 0.70.3 — streaming tool calls #4927)
• ci: remove a fat job's section containers when its run is cancelled (ci: remove a fat job's section containers when its run is cancelled #4962)
• perf(pv): cut guard-cargo's pv runtime (scoped binding lint, optimized YAML parser in dev builds) (perf(pv): cut guard-cargo's pv runtime (scoped binding lint, optimized YAML parser in dev builds) #4993)
• ci(receipts): judge receipt presence nightly over merged PRs, off the PR path (ci(receipts): judge receipt presence nightly over merged PRs, off the PR path #4999)
• test(gpu): run the six gdn device tests in the CUDA unit job again (FLAKE: intermittent SIGSEGV in libcuda during gdn device tests (aprender-gpu --lib) #4956) (test(gpu): run the six gdn device tests in the CUDA unit job again (#4956) #5044)
• release: the pass-start streak needs one night, not three (release: the pass-start streak needs one night, not three (C355 item 6) #5061)
• release: close the last strip spellings of the rehearsal name (release: the pass-start streak needs one night, not three (C355 item 6) #5061 follow-up) (release: close the last strip spellings of the rehearsal name (#5061 follow-up) #5071)
• ci: move three non-blocking PR-path rows to nightly LAB (ci: move three non-blocking PR-path rows to nightly LAB (Q6, C324) #5076)
• refactor(metrics): split davies_bouldin_score and calinski_harabasz_score under the complexity limits (refactor(metrics): split davies_bouldin_score and calinski_harabasz_score under the complexity limits #5111)
• test(metrics): pin that cluster scores ignore labels past the last row (test(metrics): pin that cluster scores ignore labels past the last row #5113)

Known issues

The known-issues entries carried from the previous release (#3715, #4590) are unchanged; see the [0.71.0] section of CHANGELOG.md in this PR.

keep-open: #3998 #3715 #3719 #4342 #4350 #4377 #4455 #4470 #4523 #4590 #4596 #4610 #4756 #4929 #4936 #4947 #4953 #4954 #5012 #5019 #5031 #5062 #5081 #5082 #5083 #5087 -- cited by the CHANGELOG for context; each closes via its own PR; the release EPIC closes at T-4

ont-delta: none -- version bump, CHANGELOG section and a carried certification; no ontology term changes

Refs: #3998

🤖 Generated with Claude Code

[noahgift]

seat 1, aprender-w71-e3a, 6dac171 — APPROVE @ 6dac171

Static review of the one commit release: 0.71.0, base 812dbe3 (the frozen main). 42 files, +9216 -210. CI on this head is the test evidence.

• Version strings.
  • Every removed or added line outside evidence/ and CHANGELOG.md was normalized (0.70.x → V on removed lines, 0.71.0 → V on added lines). The two sets are then identical: 209 × 0.70.3 → 209 × 0.71.0.
  • The single residue is apr-format = { …, version = "0.70" } → "0.71", the same bump written as a major.minor requirement.
  • No "0.70.x" string is left in any Cargo.toml at the head. No code or script changed.
• Carried crux cert (evidence/crux/0.71.0/).
  • prompt-certification.json is semantically identical to evidence/crux/0.70.3/ (jq -S), except apr_commit cc4463f → 812dbe3, the base of this bump. The whole-file diff is jq re-indentation only.
  • prompt-certification-inventory.json is byte-identical: sha256 b8e4c4626b20…, which matches the cert's inventory_sha256.
  • scripts/crux_inference_prompts.v2.json at the head hashes to cea80df5…, the cert's prompts_sha256. So prepare_bump.sh's carry precondition holds.
• CHANGELOG adds a [0.71.0] - 2026-10-11 section. I spot-read its head. I did not audit every entry against the merged PRs.

No blocking findings. NIT: the carried cert is re-indented (1 → 2 spaces), so a future text diff against the source looks like a full rewrite. A jq --indent 1 in the carry would keep the diff to the one line.

[noahgift]

APPROVE: seat 2, aprender-clerk (claude-opus-5-5, not the author), head 6dac171 (6dac171).

I read the whole diff: one commit on main 812dbe3 (= merge-base), 42 files, +9216/-210.

Version strings.

• Across every Cargo.toml and Cargo.lock, the changed lines are 209 0.70.3 → 0.71.0 pairs, plus one requirement, apr-format "0.70" → "0.71". That crate inherits the workspace version, so the requirement matches.
• No other line changes.

Measured, on a CPU host with the head checked out (git rev-parse HEAD = the head above).

• scripts/bump-version.sh --check rc 0, ending on "crates/facades/Cargo.lock matches its manifests (--locked)".
• cargo metadata --locked at the root: rc 0. All 84 path packages are at 0.71.0, and no other version appears.
• cargo metadata --locked in crates/facades: rc 0.

Carried CRUX cert (evidence/crux/0.71.0/ vs evidence/crux/0.70.3/).

• prompt-certification.json, compared key-sorted with jq -S: the only differing key is apr_commit, cc4463f7a6… → 812dbe3eb5a713f076f2e7f313c7ec6b14027a55. That is this PR's parent, and main at the time of the read. With apr_commit removed, both files hash the same. The raw text diff is large only because of key order and formatting.
• prompt-certification-inventory.json is byte-identical. Its sha256 equals the cert's inventory_sha256 (b8e4c462…).

CHANGELOG.

• One new ## [0.71.0] - 2026-10-11 section of 87 lines, inserted above the previous release.
• Repeated (#4956) and (#5062) refs are issue refs on distinct PRs, not duplicate entries.

Notes (non-blocking).

• The cert is carried, not re-run at this commit. That is as the body states: only the commit it names changed, and the prompts, manifests and admitted sets are 0.70.3's.
• Not armed by me.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=5164 head=6dac171da9076265c439007881ad439f343e53ac 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.