[noahgift]

Summary

The 0.70.2 release commit. It ships under a recorded operator emergency scope, crux-smoke, as 0.70.1 did: "0.70.2 ships on CRUX smoke on lambda and gx10 GPU. Everything bigger is nightly."

Branch release/0.70.2 is cut from main 4098007133. It holds two commits and nothing else:

• chore(release): the version bump. scripts/bump-version.sh 0.70.2, and --check is green. It touches the same 38 files as the 0.70.1 bump.
• release(0.70.2): the release content.
  • Scope entry: a crux-smoke entry for release "0.70.2" under emergency_scopes in contracts/model-capability-ladder-v1.yaml. It is data only, and pv validate is clean.
  • PP-26 witness marker: marker.json is taken from evidence(perf041): refresh the PP-26 batch-invariance witness marker #4888 and nothing else. It is the input of the declared dogfood gate scripts/check_perf041_marker.sh, which prints PASS.
  • Prompt certification: carried from 0.70.1 as byte copies. The models, prompts and oracle engines are unchanged; the release notes give the reason.
  • Release notes: evidence/release/0.70.2/RELEASE-NOTES-0.70.2.md.
  • CHANGELOG: a [0.70.2] section.

No crate source, gate or script is edited.

The notes name six known failures: #4661, #4662, #4663, #4664, #4665 and #4666. All six are on the 0.71.0 milestone. One unattended autopilot pass also moved to 0.71.

Merge back after the release, as #4776 did. Do not merge before the tag.

keep-open: #4661 #4662 #4663 #4664 #4665 #4666 -- known failures shipped in 0.70.2 and moved to 0.71.0; this release fixes none of them
ont-delta: none -- adds one emergency_scopes entry; no known-red list or ontology term changes

Test plan

• cargo fmt --all -- --check
• cargo test -p aprender-contracts --lib
• cargo test -p aprender-core --test readme_contract
• cargo deny check advisories
• pv validate contracts/model-capability-ladder-v1.yaml
• scripts/bump-version.sh --check
• scripts/check_milestone_cut.sh 0.70.2 --must-carry: PASS

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=4890 head=89e261cda1042ad14dba50034da9a9090abcdecb 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.

[noahgift]

Review by aprender-cd (non-author), verdict YES at b11fd23

This verdict covers the content of these two commits on base 4098007. It does not cover CI. At the time of review present was red and the workspace-test shards, x86-main and yoga were still pending.

Checked:

• Version bump. bash scripts/bump-version.sh --check exits 0, and the facade upstream pins and crates/facades/Cargo.lock agree on 0.70.2. Nothing in Cargo.lock is still at 0.70.1; 82 packages are at 0.70.2.
• Emergency scope. The contracts/model-capability-ladder-v1.yaml diff adds one emergency_scopes entry, crux-smoke, for release "0.70.2", with the same fields as the 0.70.1 entry (release, date, quote, hosts [lambda, gx10], thinking off/on). The quote matches the recorded operator ruling word for word. pv validate reports 0 errors and 0 warnings.
• Carried certification. evidence/crux/0.70.2/prompt-certification.json and -inventory.json have the same blob ids as evidence/crux/0.70.1/, so they are byte copies. git diff v0.70.1..tip on the certification inputs is empty: evidence/crux/0.70.1, evidence/crux/0.69.1, crux_inference_dogfood.sh, lib/crux_smoke_scope.py and crux_sweep_shards.sh. This supports the claim that the certification can be carried.
• PP-26 marker. evidence/perf041/lambda/marker.json has the same blob as evidence(perf041): refresh the PP-26 batch-invariance witness marker #4888's file. check_perf041_marker.sh prints PASS with commit=409800713, age 0.1d.
• Notes and CHANGELOG. The six known failures (0.70.2 P0: Qwen2.5-0.5B-Instruct-f16.gguf cuda serve /api/chat: gibberish output (known failure in 0.70.1) #4661 to 0.70.2 P0: qwen35-0.8b-q4km golden_output thinking_on: think block unclosed within 2048 tokens (known failure in 0.70.1) #4666) are listed with their tickets, the autopilot bar moves to 0.71 with its cause, and cells are recorded as not measured (0.70.2 P0: capability cells were not measured for 0.70.1 (~45 min/model x 26 models = 15-20 h per host) #4667). crates/ and src/ contain only version lines.

Non-blocking: in CHANGELOG [Unreleased] / Known issues, three entries still say "planned for 0.70.1" (#3715, #4590). These are stale wording, not a release gate. Fix them after the release.

Agent: aprender-cd

[noahgift]

Non-author review of #4890 at b11fd23: YES

The diff runs from merge-base 4098007133 to b11fd2381a, in two commits. be512e816a is the bump. b11fd2381a holds the scope entry, the marker, the carried certification, the notes and the CHANGELOG. Outside the bump, no file under crates/, src/ or scripts/ changes.

Measured on a detached checkout of b11fd23

check	result
bash scripts/bump-version.sh --check	rc 0
pv validate contracts/model-capability-ladder-v1.yaml	rc 0: 0 errors, 0 warnings
bash scripts/check_perf041_marker.sh	rc 0, PASS ... commit=40980071333b max_m=16 age=0.1d (window 7d)
marker.json against #4888 head c92044b3c2	byte-identical
evidence/crux/0.70.2/prompt-certification{,-inventory}.json	byte-identical to evidence/crux/0.70.1/ and to the v0.70.1 tag's copy
.prompts_sha256 against sha256(scripts/crux_inference_prompts.v2.json)	equal (cea80df5…)
.inventory_sha256 against sha256(inventory)	equal (b8e4c462…)
git diff v0.70.1 HEAD on the paths the notes name, plus scripts/crux_inference_prompts.v2.json and scripts/check_model_ladder.sh	empty

The paths the notes name are evidence/crux/0.70.1, the 0.69.1 prompt certification, crux_inference_dogfood.sh, lib/crux_smoke_scope.py and crux_sweep_shards.sh. Because that diff is empty, the notes' carry argument holds.

Read

• Scope entry. It has the same shape as the 0.70.1 entry: release, date, quote, hosts and thinking. The quote is the release owner's recorded ruling for 0.70.2, not the 0.70.1 quote with its version edited. The judge matches the whole release string, so the 0.70.1 entry admits nothing for 0.70.2 and the new entry admits nothing else.
• Carried certification and fix(serve): request tools reach the chat template; Qwen3.5 XML tool calls parse (#4650) #4866. crates/aprender-serve/src/chat_template_official.rs changed after v0.70.1 in fix(serve): request tools reach the chat template; Qwen3.5 XML tool calls parse (#4650) #4866, which added the tojson filter and request tools. The certification does not depend on that change. Its cells are oracle-engine verdicts per prompt (for example engine: llama.cpp, rows from the 0.69.1 manifests), not apr renders. Its prompt file is unchanged, and a search of the certification finds no tools or tool-call prompt. The smoke itself runs the binary built at this commit.

Advisory, not blocking

1. Stale comment. contracts/model-capability-ladder-v1.yaml still introduces emergency_scopes with the comment "OPERATOR EMERGENCY SCOPE for 0.69.1 ONLY … Refused for any other release". The list now holds three releases. Fix the wording at merge-back.
2. Notes and fix(serve): request tools reach the chat template; Qwen3.5 XML tool calls parse (#4650) #4866. The notes could name the fix(serve): request tools reach the chat template; Qwen3.5 XML tool calls parse (#4650) #4866 template change and say why it does not reach the certification, as argued above, so a later reader does not have to re-derive it.
3. present is FAILURE at this head. evidence/pr-review/4890/b11fd2381a…/ has no receipt yet. This review does not cover that check.

[noahgift]

0.70.2 state at 12:20Z: operator ruling C325 (11:45Z). A check made green by a setup fix counts only if all four hold:

• (a) the tree is clean at b11fd23 and every gate script that ran is its own;
• (b) every measured binary printed apr 0.70.2 (b11fd2381);
• (c) no expectation, threshold, prompt, oracle, budget or gate script changed;
• (d) the first red was infrastructure or setup.

Check	Counts	Why
C1 CI workspace-test	no	(d) is false: the first red was an out-of-bounds read in a test of the SIMD crate (#4891, kernel log), not infrastructure. The rerun was green but does not clear it.
C3 CRUX smoke, one of the two GPU hosts	no	(a) and (c) are false: the green needed a harness script that changes the engine order (hf and vllm first, alone on the GPU).
C4 T-1 deep	yes	The same build and smoke ran at b11, launched with APR_GIT_SHA_OVERRIDE as the nightly does. 29 of 29 binaries print 0.70.2 (b11fd2381).
C2, C3 other host, C5, C6, C7	yes	Green the first time, with nothing changed.
C8	not run

For C1 and C3 the operator is asked to choose: 2 (fix both in one new release commit; every lane reruns) or 3 (hold). Until then nothing is pushed or armed.

Agent: aprender-a3

🤖 Generated with Claude Code

[noahgift]

Replaced by #4909. This PR's head is the tagged commit 89e261cda1. The tag-push CI run on v0.70.2 attaches failed base-dependent required checks to that commit, so this PR stays blocked whatever its own run says. #4909 merges the same tree as a new merge commit (git diff 89e261cda1 fd31b82a85 is empty). The tag-push trigger is a 0.71 ticket.