fix(ci): keep per-commit range gates uncancellable

.agents/coordination.md

@@ -132,6 +132,27 @@ leaves. Owns only NEW `.agents/specs/cli-chat-complete.md`, the
 fixture, or GPU/model-download change; verification is the CPU record/doc
 checker suite. The row and open-PR list were unclaimed at selection time.
 
+**Per-commit CI gate cancellation repair (`ROAD-V1-A6`, 2026-08-01,
+`CLAIM-ROAD-V1-A6-CI-PER-COMMIT`).** Codex (GPT-5), isolated worktree
+`.worktrees/road-v1-a6-ci-per-commit`, branch
+`codex/road-v1-a6-ci-per-commit`, base `upstream/main` `1448e981`. CPU-only CI
+governance bugfix: the `documentation-checkpoint` and `commit-protocol-tag`
+jobs say their push ranges must never be cancelled, but both currently set
+job-level `cancel-in-progress: true`; a later `main` push can therefore cancel
+the only run covering the earlier push's `before..sha` range. Owns only
+`.github/workflows/ci.yml`, NEW
+`tests/scripts/test_ci_per_commit_gates.py`, NEW
+`.agents/specs/ci-per-commit-gates.md`, this claim, the `ROAD-V1-A6` roadmap
+checkpoint text, NEW
+`docs/superpowers/plans/2026-08-01-ci-per-commit-gates.md`,
+`docs/STATUS.md`, `docs/BENCHMARKS.md`,
+`.agents/parity-ledger.md`, and append-only `.agents/state.md`. No C/C++, build,
+runtime, model, kernel, fixture, README, lifecycle, support, or benchmark change;
+verification is the CPU Python mutation test plus all record/document checkers.
+No GPU, model download, compiler, external host, or parallel agent is required.
+**LANDED + CPU-GATED locally:** focused invariant 3/3 and full script suite
+98/98; push/PR and hosted CI are pending.
+
 **Canonical DONE-owner reachability repair (`KV-PREFIX-CACHE`,
 `SAMPLE-LOGPROBS`, `SPEC-DFLASH`,
 `MODEL-SPEC-qwen3-dflash-dflash-qwen3-for-causal-lm`,

.agents/parity-ledger.md

@@ -886,3 +886,4 @@ Columns:
 | 2026-07-31 (`SERVE-C-ABI` W0 contract spike; `CLAIM-SERVE-C-ABI-SPIKE`; CPU-only records/docs) | Accepted `.agents/specs/c-api-library.md` for the already-shipped original C packaging layer: complete scope, vLLM semantic chain/deviation, ABI v10/19-symbol baseline, ownership/error/version/dispatch rules, exact code/test anchors, gates, dependencies, risks, and W1-W5 follow-ons. Also fixes the verified stale public `VLLM_ABI_VERSION 9` labels in README/USAGE to the source-of-truth v10 and adds the missing v10 usage-table entry. No production/test/CMake source changed. | Pinned vLLM `555967922` has no C ABI; behavior beneath the adapter remains owned by its vLLM-derived engine rows. The flat ABI is the recorded llama.cpp-style packaging deviation and may translate, never reimplement, policy. | **CPU/records gate only; benchmark NOT APPLICABLE.** Focused C11/C++/dlopen/export gate passed 3/3 after explicitly building `vllm_shared`; five record checkers pass. `check-agent-record` reports the base tree's same six missing closing-commit objects (`444ea9d7`, `7a3f04b2`, `164453a2`), none in this row/diff. Row stays `ANCHOR-BACKFILL` because all-symbol dlsym coverage (chat symbols currently omitted), historical-layout compatibility, allocation-failure no-throw proof, lifetime sanitizer stress, and a standalone real-model C consumer remain W1-W5. |
 | 2026-07-31 (`CLAIM-CPU-GCC12-WERROR-PORTABILITY`; maintenance, rows `QUANT-GGUF-KEEPQ-LOADER` + `KV-OFFLOAD`; lifecycle unchanged) | Removes two GCC 12 production-library `-Werror` blockers without suppressions: the GGUF prefault keeps the same one-byte-per-page volatile XOR but uses simple assignment, and the KV filesystem tier builds the identical `.<pid>.<counter>.tmp` suffix with append operations inside its thread-local initializer. No API, algorithm, default, CUDA, fixture, or golden change. | Behavior remains grounded in the accepted loader and KV-persistence leaf specs: llama.cpp mmap prefault intent and vLLM `tiering/fs/io.py` unique temporary-file publication. This is compiler portability, not a parity-surface change. | RED: GCC 12 failed first at `qwen3_5_gguf_weights.cpp:49` (`-Wvolatile`), then at `fs_io.cpp:66` (`-Wrestrict`). GREEN: production `vllm` and focused test targets build clean; focused CTest 2/2 (`test_gguf_keep_quant`, `test_kv_offload_fs`). Full all-target build is PARTIAL at 42% on unrelated test-only GCC 12 `-Wrestrict` diagnostics in `test_deepseek_v2_paged_engine.cpp` and `test_glm4_moe_lite_paged_engine.cpp`; no full-CTest claim. Benchmark NOT APPLICABLE. |
 | 2026-08-01 (`SERVE-CLI-CHAT` W0 contract spike; `CLAIM-SERVE-CLI-CHAT-SPIKE`; CPU-only records/spec) | Accepts `.agents/specs/cli-chat-complete.md`, corrects the inventory from “no direct commands” to the actual pinned `chat`/`complete` surface, and decomposes a dual-mode port: exact remote OpenAI HTTP/SSE commands plus preservation of the existing in-process invocation as a compatibility alias. No production, test, CMake, model, kernel, fixture, or generated file changes. | Pinned vLLM `5559679229`: command registration `vllm/entrypoints/cli/main.py:17-37,73-98`; model/auth resolution and stream shaping `vllm/entrypoints/cli/openai.py:30-100`; chat `:155-234`; complete `:237-312`. The local compatibility baseline is `examples/cli/main.cpp:1-207`. | CPU record/doc gates only; benchmark `NOT APPLICABLE`, `benchmark_binding=false`. Implementation remains absent and the row moves `INVENTORIED` -> `SPIKE`. W1-W5 name parse, transport, complete, chat, and packaging gates, including fake-server request/SSE transcript parity, Release `-Werror`, ASan+UBSan, and TSan. |
+| 2026-08-01 (`ROAD-V1-A6` per-commit gate cancellation repair; `CLAIM-ROAD-V1-A6-CI-PER-COMMIT`) | Removes job-level cancellation from the documentation-checkpoint and commit-protocol-tag jobs so every push's unique `before..sha` range receives a verdict. Adds a standard-library static invariant and mutations for both forbidden keys on both protected jobs. No runtime, lifecycle, support, model, build, or benchmark behavior changes. | No vLLM runtime counterpart; this enforces the local `AGENTS.md` commit/checkpoint protocol and the range semantics in `.github/workflows/ci.yml`. | RED: current workflow produced four violations. GREEN: focused 3/3 and full script suite 98/98. Benchmark NOT APPLICABLE (`benchmark_binding=false`); CPU-only governance evidence. |

.agents/roadmap_v1.md

@@ -535,6 +535,12 @@ diagnostics recorded in `docs/BENCHMARKS.md`.
 | `SERVE-E2E-NIGHTLY` (formerly A5) | e2e suites per gates.md (server conformance nightly on dgx etc.) | ☐ next; leaf spike required |
 | A6 | **User-facing surface closure** (audit `CLAIM-DOCS-COVERAGE`, 2026-07-24): make what SHIPPED reachable and truthfully documented. The audit found 2 class-D overstatements, 5 stale doc lines, 4 shipped-but-unreachable features and 153 undocumented env vars | 🚧 Tier 1 (falsehood repair) DONE in the audit commit. **Tier 2 DONE 2026-07-24 (`CLAIM-DOCS-T2-FIXES`)**, 3 of its 4 items: the D1 guard (a per-connector worker-half capability predicate; the disk connector is now REFUSED at engine construction on every device instead of silently shortcutting prefill for KV nothing writes — **its worker half remains unimplemented and is NOT claimed**), server `--tool-call-parser`/`--reasoning-parser` (40 tool names / 7 reasoning parsers now reachable over HTTP; defaults reproduce the old hardcode exactly), and server `--kv-transfer-config` mirroring vLLM's own CLI + JSON shape. Also landed from Tier 3/4: `docs/KV-OFFLOAD.md`, the README flag rows, the three missing Gemma model rows, and the `SERVE-C-ABI` 17→19 record repair. **Comprehensive refresh DONE 2026-07-25 (`CLAIM-DOCS-REFRESH`)**: **Tier 4** `docs/ENVIRONMENT.md` (26 user-facing env vars documented of 153 swept, kernel-internal tail on `scripts/env-doc-allowlist.txt`) + its CI checker `scripts/check-env-doc.py` (mutation-tested, fails on a new undocumented+unallowlisted production env var); the **Tier 3** README refresh to current git history (spec-decode `--speculative-config` now documented as shipped, the "not user-visible" class-B falsehood fixed, C ABI `VLLM_ABI_VERSION 6`, CLI spec-config flag); new usage guide `docs/SPECULATIVE-DECODING.md`. STILL OPEN: `vllm-cli` structured-output/sampling flags (Tier 2 item 4, code, not docs). **C-ABI contract W0 2026-07-31 (`CLAIM-SERVE-C-ABI-SPIKE`):** accepted [c-api-library.md](specs/c-api-library.md), corrected public ABI v9 labels to v10, and named W1-W5 compatibility/no-throw/lifetime gates; `SERVE-C-ABI` honestly remains `ANCHOR-BACKFILL`. **CLI chat/complete W0 2026-08-01 (`CLAIM-SERVE-CLI-CHAT-SPIKE`):** [cli-chat-complete.md](specs/cli-chat-complete.md) corrects the inventory (the pin does ship both remote commands), selects exact OpenAI HTTP/SSE parity plus a compatibility-preserved local invocation, and decomposes CPU-only W1-W5 implementation leaves; `SERVE-CLI-CHAT` is `SPIKE`, no command code has landed. Full table: [specs/docs-coverage-audit.md](specs/docs-coverage-audit.md) |
 
+`ROAD-V1-A6` CI-governance checkpoint (2026-08-01):
+`CLAIM-ROAD-V1-A6-CI-PER-COMMIT` completed the CPU-only
+[per-commit gate cancellation repair](specs/ci-per-commit-gates.md). Both
+unique-range jobs are uncancellable and mutation-gated; no runtime capability
+or lifecycle state changes.
+
 ## B. Research tracks (complete)
 
 The B1-B7 parallel research block closed on 2026-07-10. Its frozen questions,

.agents/specs/ci-per-commit-gates.md

@@ -0,0 +1,76 @@
+# Per-commit CI gate cancellation repair
+
+## Scope
+
+This spike owns `ROAD-V1-A6` and only the two diff-scoped CI jobs
+`documentation-checkpoint` and `commit-protocol-tag`. It removes their ability
+to cancel an earlier run and adds a CPU static regression test. It does not
+change runtime code, lifecycle states, support claims, build jobs, or the
+workflow-level pull-request deduplication policy.
+
+## Upstream chain
+
+There is no vLLM runtime counterpart. This is repository governance around the
+local contracts in `AGENTS.md:80-91,230-245` and the range selection in
+`.github/workflows/ci.yml:73-135`. GitHub Actions evaluates job concurrency by
+group and cancels an in-progress member when `cancel-in-progress` is true. The
+two jobs group by `github.ref`, but each push checks only
+`github.event.before..github.sha`; a replacement run therefore cannot cover a
+cancelled predecessor's unique range.
+
+## Our baseline
+
+`.github/workflows/ci.yml:74-77,87-99` simultaneously says the two gates must
+not be cancelled and configures both with `cancel-in-progress: true`. Existing
+Python script tests validate checker semantics, but no test pins the workflow's
+non-cancellation invariant.
+
+## Port map
+
+- `.github/workflows/ci.yml`: delete only the two job-level concurrency blocks.
+- `tests/scripts/test_ci_per_commit_gates.py`: parse the workflow text by job
+  indentation and reject `concurrency` or `cancel-in-progress` inside either
+  protected job; mutation-test that each forbidden key is caught.
+
+No YAML dependency is added. Text inspection is deliberate because the test
+guards the exact workflow structure and runs in the standard-library suite.
+
+## Tests to port
+
+No upstream vLLM test applies. The local executable contract has three cases:
+the repository workflow passes; injecting a concurrency group under either
+protected job fails; injecting `cancel-in-progress: true` under either protected
+job fails. The test must be RED against the current workflow before the fix.
+
+## Gates
+
+- `python3 tests/scripts/test_ci_per_commit_gates.py`
+- `python3 -m unittest discover -s tests/scripts -p 'test_*.py'`
+- all repository record, documentation, structure, fusion, leakage, and env
+  checkers
+- `git diff --check`
+
+Correctness is the static workflow invariant. E2E is the mutation suite.
+Performance and memory are `NOT APPLICABLE`; no executable runtime path changes.
+GPU/backend gates are `NOT APPLICABLE`; all verification is CPU-only.
+
+## Dependencies
+
+Depends only on Python 3, Git, and the checked-in workflow. No model, compiler,
+GPU, external service, dependency source, license addition, or download is
+required. The row was unclaimed at selection time and does not overlap open PRs.
+
+## Work breakdown
+
+1. `CI-PC1`: add the static test and observe RED on both protected jobs.
+2. `CI-PC2`: remove their job-level cancellation blocks and observe GREEN.
+3. `CI-PC3`: refresh checkpoint records, run the full CPU governance suite,
+   commit, push, and open the PR.
+
+## Risks/decisions
+
+The chosen design removes cancellation only from the two range-scoped jobs.
+Removing all concurrency would waste build capacity; changing the push range to
+recheck history would duplicate work and complicate new-branch handling. Pull
+request workflow-level cancellation remains safe because the newest PR run
+checks the full PR base-to-head range.

.agents/state.md

@@ -34505,3 +34505,21 @@ The required Slack selection notification was attempted through the bundled
 secret-safe sender to the only conventional target available, `#general`, but
 Slack returned `channel_not_found`. No channel ID/name is configured and no
 credential was inspected or exposed.
+
+## 2026-08-01 — `ROAD-V1-A6` per-commit CI cancellation spike
+
+`CLAIM-ROAD-V1-A6-CI-PER-COMMIT` selected an unclaimed CPU-only governance bug.
+The documentation and commit-protocol jobs each validate one push's unique
+`before..sha` range but are grouped by ref with cancellation enabled, so a later
+main push can permanently suppress the earlier range. The accepted spec limits
+the fix to removing those two job-level cancellation blocks and adding a
+mutation-tested static invariant. No implementation, runtime, lifecycle, GPU,
+model, or benchmark change is claimed at this checkpoint.
+
+## 2026-08-01 — `ROAD-V1-A6` per-commit CI cancellation repair implemented
+
+`CI-PC1` RED produced four violations, one `concurrency` and one
+`cancel-in-progress` finding for each protected job. `CI-PC2` removed only those
+two job-level blocks; workflow-level PR deduplication and every tree-scoped job
+remain unchanged. Focused test is 3/3 and full script unittest discovery is
+98/98. Push, PR, hosted CI inspection, and the second Slack notification remain.

.github/workflows/ci.yml

@@ -56,9 +56,6 @@ jobs:
           python3 scripts/check-runner-routing-consistency.py
           python3 tests/scripts/test_check_runner_routing_consistency.py
   documentation-checkpoint:
-    concurrency:
-      group: ci-doc-${{ github.ref }}-${{ github.repository }}
-      cancel-in-progress: true
     # Gate: every code/test/benchmark/spike/lifecycle iteration refreshes both
     # user-facing status surfaces in that same commit, including void attempts.
     #
@@ -88,9 +85,6 @@ jobs:
           fi
           python3 scripts/check-doc-checkpoint.py --base "$base" --head "$head"
   commit-protocol-tag:
-    concurrency:
-      group: ci-commit-${{ github.ref }}-${{ github.repository }}
-      cancel-in-progress: true
     # Gate: every NEW commit must carry the FOLLOWING_AGENTS_PROTOCOL trailer,
     # asserting the contributor read AGENTS.md. See .agents/ai-coding-assistants.md.
     #

docs/BENCHMARKS.md

@@ -42,6 +42,13 @@ Two byte-exact fronts on the GB10 NVFP4 decode graph (`~/laguna-xs-nvfp4`, ids `
 
 The 78 removed casts are ~78 µs/step — BELOW the ~0.3 ms nsys run-to-run noise (the dominant `enable_if` projection GEMV drifts ±0.4–0.7% between runs), so GPU-busy reads at parity; the win is the deterministic node-count drop (fewer graph nodes → less ramp/drain the CUDA graph doesn't hide), reflected as a neutral-to-+0.29% `decode_hp`. Lands default-ON on the same "shrink the captured node count, byte-exact" basis as the glue/preamble/addnorm folds (`=0` is a same-binary A/B opt-out). The remaining decode tail (projection GEMV ~55%, lm_head 7.4%, we-win Marlin MoE ~12%) is at/beyond parity; the load-time Marlin repack (`TransposeToInt32`/`gptq_marlin_repack`/`ProcessScales`, 20046 inst) is a one-time cost, correctly zeroed by the 2-length diff (the nsys-aggregate trap).
 
+## CI per-commit gate repair (2026-08-01, `CLAIM-ROAD-V1-A6-CI-PER-COMMIT`) - NOT APPLICABLE
+
+The workflow-only repair changes no runtime or performance path
+(`benchmark_binding=false`). Reproduction:
+`python3 tests/scripts/test_ci_per_commit_gates.py` (3/3) and the full script
+suite (98/98).
+
 ## GCC 12 production-library portability (2026-07-31, `CLAIM-CPU-GCC12-WERROR-PORTABILITY`) - NOT APPLICABLE / all-target build PARTIAL
 
 This maintenance checkpoint changes no runtime algorithm or benchmark axis.

docs/STATUS.md

@@ -1,5 +1,10 @@
 # vllm.cpp status
 
+**CI governance checkpoint (2026-08-01):** `ROAD-V1-A6` per-commit cancellation
+repair is implementation-complete and CPU-gated. The documentation and
+commit-protocol jobs cannot cancel their unique push ranges; no runtime
+capability or support state changes.
+
 **Roadmap-v1 reality audit (2026-07-31, `CLAIM-ROADMAP-V1-AUDIT`, `.agents/specs/roadmap-v1-audit-2026-07-31.md`):** an 8-lane code-grounded audit + adversarial DONE-verify (26 agents) vs the 18-row plan-of-record. Correctness + CPU-gated features are REAL; the dominant open gate across nearly all RI rows is **every-axis SPEED**; most GPU "DONE" proofs are DGX-recorded (not re-verifiable on the current box). **4 DONE claims refuted:** MM-video is near-tie not STRICT; Gemma-4 image is 16/18 near-tie not STRICT 18/18 (audio unbuilt); ROAD-V1-A (async code DONE but both SGLang-floor arms never ran, 27B 114/124 coin-flip not clean); C6 async-serving (SERVE-ASYNC-LLM still GATING). D5 LoRA is an unwired standalone CPU brick. Fixed `backend-matrix` BACKEND-GATE-CUDA-VLLM staleness 49/124 → 114/124. (Follow-up: reworded a spec code-snippet `check-agent-record` misread as a dangling link — rc=0.)
 
 This is the per-capability status ledger. [README.md](../README.md) is the

docs/superpowers/plans/2026-08-01-ci-per-commit-gates.md

@@ -0,0 +1,58 @@
+# Per-commit CI Gates Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Ensure the documentation and commit-protocol range gates can never be cancelled before validating their unique push range.
+
+**Architecture:** Keep the existing range computation unchanged and remove job-level cancellation from only the two protected jobs. Add a standard-library Python test that extracts those job blocks and mutation-proves both forbidden concurrency keys are rejected.
+
+**Tech Stack:** GitHub Actions YAML, Python 3 `unittest`, repository record checkers.
+
+## Global Constraints
+
+- CPU-only, with no GPU, model, compiler, service, or download.
+- Do not change workflow-level PR deduplication or any tree-scoped job.
+- Do not change runtime code, lifecycle state, support claims, or benchmarks.
+
+---
+
+### Task 1: Pin the non-cancellation invariant
+
+**Files:**
+- Create: `tests/scripts/test_ci_per_commit_gates.py`
+- Modify: `.github/workflows/ci.yml`
+
+**Interfaces:**
+- Consumes: the `jobs:` mapping and the two job names in `.github/workflows/ci.yml`.
+- Produces: `job_block(text: str, name: str) -> str` and `gate_errors(text: str) -> list[str]` in the test module.
+
+- [ ] **Step 1: Write the failing test**
+
+  Add tests asserting the checked-in workflow has no `concurrency:` or
+  `cancel-in-progress:` key inside either protected job, plus mutations that
+  inject each forbidden key and assert an error names the job and key.
+
+- [ ] **Step 2: Run test to verify it fails**
+
+  Run: `python3 tests/scripts/test_ci_per_commit_gates.py`
+
+  Expected: FAIL for both repository job blocks because each currently contains
+  `concurrency` and `cancel-in-progress`.
+
+- [ ] **Step 3: Write minimal implementation**
+
+  Delete only the four-line job-level concurrency blocks beneath
+  `documentation-checkpoint` and `commit-protocol-tag`. Retain their comments,
+  range selection, and workflow-level concurrency unchanged.
+
+- [ ] **Step 4: Run test to verify it passes**
+
+  Run: `python3 tests/scripts/test_ci_per_commit_gates.py`
+
+  Expected: PASS, including both mutation families.
+
+- [ ] **Step 5: Run full verification and commit**
+
+  Run the script unittest discovery, all repository governance checkers, and
+  `git diff --check`; update the owned checkpoint records; commit with
+  `FOLLOWING_AGENTS_PROTOCOL` and `Assisted-by: Codex:GPT-5 [Codex]`.