M28g implementation and 1.2.0 release record

This chronological record begins with the 1.2.0 source candidate rebased onto GitHub main at 2d8ce6a9f5627b0bb16ffd6c4c9b0619f1d2c749 and ends with the owner-selected release evidence below. Intermediate statements retain their original checkpoint date and scope. The 1.1.0-beta release record remains at its original scope.

Task record

Title/ID: m28g-macos-native-installer, m28g-ubuntu-arm64-native-installer, m28g-host-clients-as-built-alignment, m28g-installation-and-integrated-adoption, m28g-release-b-qualification, m28g-mcp-tool-catalog, m28g-a2a-agent-facade, m28g-integration-live, m28g-pypi-1.2.0b0-publication.

Milestone: 28g — Installation, Integrated User Qualification and Release B.

Goal: Build one version-bound host candidate and qualify native installation, cross-client operation and the assembled release on every selected profile.

Inputs: M28–M28f source and component records; selected compiler inventory and host integration matrix; 1.2.0 release source; local Mac, Linux ARM64 and Pi 4 test-bed descriptions. Private signing identities, publisher trust and target credentials remain outside this record.

Changes:

• Native Mac and Ubuntu ARM64 package builders stage only an independently pinned archive, keep guest artifacts optional, and publish output only after package signing or notarization succeeds. Their staged code permissions allow ordinary users to read and launch installed tools, with the copied Mac app signature rechecked after permission normalization.
• The compiler-selected host archives now include Cohesix Python and NeMo kit wheels at 1.2.0b0, the PEP 440 spelling of Release B 1.2.0-beta, plus the NeMo lock, installer and digest report. Both wheel versions must match the selected release before an installer stages them.
• SwarmUI resolves installed tools from the fixed OS location and stages a native Mac icon; app and extension version metadata identify 1.2.0.
• coh doctor reports the selected MCP/A2A switches without claiming a live endpoint probe. SwarmUI help fixtures follow the implemented console grammar.
• Public candidate, package and Python instructions identify 1.1.0-beta as the published release and keep 1.2.0-beta qualification explicit.

Commands: scripts/check-generated.sh; focused installer, release, Python and NeMo pytest files; cargo fmt --all -- --check; cargo check --workspace; cargo test --workspace; cargo clippy --workspace --all-targets -- -D warnings; cargo audit; cargo deny check advisories; scripts/ci/test_plan_run.sh --list.

Checks: Generated inventory, selected wheel bytes, source/manifest membership, installer path and byte bounds, failed-signing output isolation, Rust and Python contract tests, and exact proof-class separation. Package staging and synthetic tests cannot establish installed-file, desktop, target or release acceptance.

Deliverables: Source-level candidate and focused test logs retained in ignored out/m28g/. There is no passing native installer qualification, live adoption case, physical Pi release qualification, integrated Test Plan run or release-owner approval. A later Pi read/protocol diagnostic is recorded below at its own source and image identity.

Current validation and blockers

The focused Python installer, release bundle, release qualification, package and provider-matrix tests passed 96 cases after rebasing; generated consistency, Test Plan catalog, Rust formatting, workspace check and tests, audit and advisory checks passed on the rebased candidate. Focused coh doctor and SwarmUI help tests passed after correcting stale selected contract assertions. The rebased Rust suite output is retained in ignored out/m28g/cargo-test-rebased.log. No exact target or installed-package claim follows from it.

The release-version wheel alignment passed 47 focused NeMo, Python, package and bundle tests. scripts/check-generated.sh passed after regenerating the compiler inventory, and both built wheel METADATA records report 1.2.0b0. The fresh QEMU seL4 production profile in this isolated checkout passed release-mode source/artifact validation. A QEMU image built at predecessor commit 616d34b86eab booted four cores, reached the serial console and answered ping with PONG. The rootserver’s [BUILD] marker matched that commit. This is boot liveness only, without authenticated TCP, live job or final-source release evidence.

The installed Rust toolchain exposed Clippy warnings in the rebased source. The candidate now repairs the large ledger enum and argument grouping without changing admission, job or compiler semantics; it also removes narrow style warnings. Full cargo clippy --workspace --all-targets -- -D warnings passed, along with focused ledger, gateway, GPU child and compiler tests. No lint suppression was added.

The first Stage 01 retry passed generated consistency, formatting, Clippy, workspace check/tests, SwarmUI, coh, QEMU and Pi feature tests, Pi runtime tests and the AArch64 target check. Its Python action failed 35 cases: 33 needed the isolated checkout’s repository .venv, one exposed the existing QEMU/Pi gateway-control difference in the parity test, and one found the wheel inspector’s stale optional-extra list. After provisioning the pinned .venv and repairing those test contracts, the focused 85-case set passed. The next Stage 01 attempt passed every preceding action and reached 2,913 passing Python tests; three release-bundle fixture cases still used the old wheel metadata. The fixture now describes the selected 1.2.0b0 wheel and the affected nine focused tests passed. On clean commit bf0fd4bda, Stage 01 passed completely, including 2,916 Python tests, 116 Python subtests and the Rust risk ratchet. Provisioned QEMU Stage 02 passed against the same source digest after the Test Plan correctly required a fresh attestation for the selected private Queen credential and manifest copies.

QEMU Stage 03 then built separate base/gated exact-source images and ran all 19 authenticated TCP regression scripts successfully. Its aggregate and both artifact verifiers passed, but the stage refused final PASS when generated MCP and A2A catalogues differed from their initial contents. The batch runner’s generated-output snapshot inventory omitted those two newer files. The inventory now includes both catalogues, and all 21 focused batch-wrapper tests plus scripts/check-generated.sh pass after restoring the committed generated outputs. The 19-script attempt remains a retained transport result, not accepted Stage 03 evidence; the corrected source requires fresh stages.

At clean commit 522463fa1ade7fcf01092ab4754825fe71dcf5dd (source digest sha256:a730a325a7055c5614e9b93dffc6342db67462046c3183c26e2c1283621dfc73), the canonical QEMU and physical Pi 4 plans each passed Stages 01–05. The QEMU Stage 05 record is under ignored out/m28g/full-plan-522-qemu; the Pi record is under out/m28g/full-plan-522-pi4-v3. Both due-diligence gates passed with the owner’s renewed EX-2026-0024 exception. The earlier failed Stage 05 attempts remain failed at their original identities. Pi runtime/DMA acceptance uses same-boot UART and authenticated Queen-log captures, whose separate hashes and composition are recorded in pi4-stage5-proof-provenance.json; it is not one uninterrupted serial capture. These staged results do not qualify a native installer, Jetson KVM pressure, packaged SD image or assembled Release B.

The 020a516300 to 522463fa1ade change contains only BUILD_PLAN, STATUS, HOST_TOOLS and audit prose, two operator skills, and the renewed exception register/finding. The tracked runtime, manifest, generated contracts, pressure runner and benchmark workload are byte-identical across those commits. For the Mac HVF pressure contract, this is a nonmaterial source change: retain the original medium and high results under out/m28g-pressure-020a-r1/ in the pressure checkout, with summary SHA-256 values 130403cc91d43e064d521cb467437782671cd1621b9bdcb5632a89677fd2cd92 and 8ccbb60a5678a040bebe098854d6c1321e9d4f6da38d9abc9e34814c279f7cef. This equivalence decision does not rename those results as 522463fa1ade tests, supply Jetson KVM pressure or apply automatically to later code changes.

During the M28g Jetson KVM preflight, the M26e pressure runner’s process-name probe matched the runner or its launching shell when the pinned QEMU path or Cargo path appeared in its own arguments. The canonical check stopped before measurement; selecting the same binary through a short private alias proved that the original false match was in process scanning. A separate attempt correctly refused a live log descriptor under out/; its rerun writes the live log outside the checkout. Under m28g-kvm-pressure-runner-portability, the runner now exempts only its own process ancestry and still checks unrelated processes, open output writers and ports. Bash syntax and 60 focused pressure tests passed. The Jetson replay at source 522463fa1ade passed authenticated NineDoor target operation, three service fault injections, critical duties and three Worker fault injections. It then failed before medium-load acceptance: host integration selected macos-arm64 on the Linux AArch64 KVM host. That run remains a failed KVM acceptance attempt; it cannot qualify this repair or a later release source. The selected Linux AArch64 matrix now includes only the three target-runtime rows, and the runner passes linux-aarch64 explicitly. The focused host-profile and pressure CLI tests and a fresh exact-source KVM replay are required before this restoration task can close.

Six pre-existing Jetson user services were stopped for the KVM quiescence preflight. Three GPU session services and the original LoRA reference service restarted. Two older LoRA unit files referenced an absent source checkout; their service processes therefore failed on restart. The hf_native.py file at that path was restored byte-for-byte from the M28b implementation commit 8c0147bad1c7dc893bac84d791da0166ec3d2c48 (SHA-256 b4504c4807f9fc7c3af0d9f0e8981cdf9bc1d110ee16ba45cbcc5040aa6855e7). Both units then reported active/running. This is test-bed restoration, not new provider or release acceptance evidence.

On 27 September 2026, Apple issued the G2 Developer ID Installer certificate CJ48JZP37S for team KB88FQXUX2, valid through 17 September 2031. Its public key matches the locally generated CSR; the private key stays in the login keychain. security find-identity lists the installed identity 5F2F0CCE752EE22658176BBE0A99111F03878A8C. A disposable probe package passed productsign and pkgutil --check-signature with Apple’s trusted timestamp, and the stored cohesix-m28c notarytool profile authenticated to the same team. These checks establish usable signing credentials, not a signed Release B installer or installed lifecycle acceptance. The available Linux ARM64 hosts have no selected publisher signing key or public Debian maintainer identity. Ubuntu 22.04 package candidates were observed in a Jammy chroot and Ubuntu 24.04 packages on a native host; neither is a clean signed .deb installation. No Ubuntu 26.04 ARM64 installation host has been selected. The selected Python and NeMo wheels are source candidates; there is no final exact 1.2 archive or installed cross-client parity report.

The source candidate now has an installer qualifier for independently trusted publisher signatures, package-manager receipts and installed byte readback. It has not run on a signed package or installed host and does not cover GUI launch or package lifecycle. The Release B archive verifier now requires both native installer results at the same source and archive hashes. The conformance matrix still has no m28g-adoption-live or m28g-integration-live case runner. The required capability-by-capability installed-client parity matrix and frozen numeric adoption, quality, recovery and control-latency budgets have not been written. These source and evidence gaps remain even after signing identities and a Ubuntu 26.04 test host become available.

The selected QEMU and Pi manifests now both enable the host MCP/A2A protocol ceiling. This selection changes the Pi host profile and its generated Python contract and target fixture. The schema remains disabled when a switch is absent, and the compiler still requires the master and each protocol switch to be true. The production-profile test guards the selected values. The earlier Pi-selected diagnostic run used a separate source candidate with both protocols disabled and stopped after approximately eight minutes at the owner’s request. Its authenticated TCP and read-only observations remain diagnostic evidence only. The later enabled Pi run below supplies a fresh exact image, matching host tools and empty MCP/A2A discovery. It does not supply native protocol jobs or the M28g integrated matrix required for release claims.

At the earlier beta-source checkpoint, the A2A Agent Card advertised 1.2.0-beta and both Python wheels carried 1.2.0b0. The later stable-version change below supersedes these packaging values; the earlier observations remain bound to their original source.

M28g protocol-mode deployment repair

Discovery task: m28g-release-b-qualification. Restoration task: m28g-integration-live. The selected release builds MCP and A2A on, but the host gateway previously rejected every launch setting for those routes. A single installed gateway could therefore serve only the both-enabled mode; the required MCP-only, A2A-only and neither modes needed different compiled catalogues. The gateway now validates both generated catalogues against their selected manifest before accepting the exact launch value false for the master, MCP or A2A switch. Each value only closes a route for that process; other values fail startup, and generated-disabled routes cannot be opened. The focused live-process test starts the actual gateway binary in all four modes, checks its route statuses and retained authenticated REST route, and checks malformed or widening launch values. Three focused tests passed at the working source. These use a mock Queen and prove gateway routing only; they are not m28g-integration-live native-target, job, revocation or recovery evidence. Host CLI, Python and benchmark clients retain the same endpoint contracts and original identities; only gateway route availability narrows.

Under m28g-installation-and-integrated-adoption, the conformance matrix and Test Plan now select a bounded m28g-adoption-live verifier. It binds the assembled release qualification, both native installer results and a named evaluator’s raw, hashed walkthrough at one source commit. Lukas Bower may evaluate the sole-developer release as owner-evaluated, without an independence claim. The walkthrough records the three journeys, native desktop launch, doctor, data-preserving rollback/uninstall and per-host steps, time, core downloads and interventions. The numerical limits were frozen in BENCHMARKS before any integrated case: the comparable 1.1.0-beta archive baselines are 26,501,100 Mac and 28,513,510 Linux bytes. The new adoption verifier reports live_host only and explicitly leaves native target outcome reverification to the owning provider cases. Focused matrix, parser and catalog checks passed. At this checkpoint, no evaluator walkthrough or native Release B installer result existed; later receipts are recorded below, while adoption remains open.

The m28g-integration-live matrix case now starts the exact installed host gateway against one private-address Queen in all four effective modes. It requires the same authenticated boot manifest and terminal result bytes after each process restart. Its retained ordinary MCP and A2A SDK records must share the original native admission and independently checked CUDA terminal; the separate subject, revocation, budget, Queen-loss and latency observations must bind their raw files and stay inside the frozen BENCHMARKS limits. The test runner rejects public target addresses and passes only credential references to the gateway launch. Focused mode-selection, identity, malformed input and latency tests use controlled host fixtures. This is an executable acceptance contract, not a passing live case: installed 1.2.0 bundles, a final-source Queen, native result and external observations remain to be supplied and run.

M28g also still requires frozen adoption and quality budgets, a named, raw-backed clean installation evaluation, the four effective protocol modes, native CUDA, PEFT, Apple and NeMo journeys from installed bytes, packaged QEMU, physical Pi installation evidence, and assembled release due diligence. These gates must bind one final source and package identity before the milestone can be marked Complete. Material AI assistance produced this implementation checkpoint; every claim above is limited to the executed source and host checks.

During the requested 60-minute operational burn-in setup, the exact Pi image builder refused the d38ab5a4f source after Pi-profile code generation changed the MCP and A2A catalogues. Its generated-output allowlist omitted both files. The builder now classifies those compiler-owned outputs correctly while still rejecting unregistered changes, and the focused guard test passed. This repair does not make the in-progress burn-in, Pi image, or release qualification pass.

The first protocol-enabled Pi candidate at 1870cca48 booted from verified RAM bytes and its first authenticated TCP read returned the selected manifest hash. The Mac gateway connected, but MCP job discovery and the A2A Agent Card returned 403 because no native standing ledger was configured on that host. This is a deployment gap, not a Pi protocol-switch failure. The gateway now answers authenticated discovery with no job tools, templates or skills when standing authority is absent; it continues to refuse job admission. A focused gateway integration test covers both protocol projections. This repair needs a new exact-source image and host binary for the replacement burn-in; the earlier boot is diagnostic evidence for its own source only.

Direct host compute alongside the Pi diagnostic

At source 05e1facdd008240a12dfe133c033912dcc87d5f5, the selected 1.2.0b0 NeMo kit wheel was installed without editable source on the Jetson Orin Nano (sm_87, driver 595.78). Its wheel SHA-256 is 91535a61c428021fe0167087fa8a6e7a5cfddd857ee56bb225364b4bd293f7e2; the pinned dependency lock and pip check passed. The maintained CUDA reference runner built the native helper and bridge from that source and passed real vector addition, matrix multiplication, wrong-device, over-budget, stale-inventory, timeout and cancel cases. The helper SHA-256 is 495eb3cd0a8c0591a1c03e59dd32ef3aff2adbf8a29c1ad3c0033a930722271d; the bridge SHA-256 is 72780ce2aff934dbef6eead8ab78f10d4c2b5fe8de94bbb65b4ea5a3cd5de6f4. The maintained summary explicitly marks this seven-case reference result non-authoritative, without a Worker receipt or admitted workload transport.

On the Mac’s Apple M4, pinned MLX 0.32.2 ran a real Metal matrix product and matched every element against an independent NumPy result. The 1.5B Qwen2.5 four-bit model at revision 8b403126fc14f14cfc99bb4cfa72ecbc129ea677 was copied to ordinary local files and bound by model-tree SHA-256 c0a2253c519413b712780076c9188676bd006859bfa55bf3acb5818cd3274927. The bounded Cohesix MLX service generated the expected one-character answer to a fixed arithmetic question on Apple M4 Metal, with output SHA-256 4b227777d4dd1fc61c6f884f48641d02b4d121d3fd328cb08b5531fcacdabf8a and observed 967,437,720 bytes peak Metal allocation.

The Jetson’s NeMo Toolkit 1.9.0 then ran its native direct evaluation against the same Mac model through a protected reverse SSH tunnel and MLX-LM’s local OpenAI-compatible server. Its native profiler recorded WORKFLOW_START, LLM_START, LLM_END and WORKFLOW_END, no error, and the expected answer; the private evaluation log SHA-256 is 2104745e1f67d09ce19fe22aa795d6f476e36bccae467c750a751f07e2786a2f. An earlier diagnostic request to Cohesix’s deliberately bounded MLX service was refused because NeMo sent system and user messages while omitting explicit token and temperature fields; its native eval process still exited zero, but the profiler had no LLM_END. That attempt is failed compatibility evidence, not a successful model evaluation. The second server supports NeMo’s request shape for this direct host check. Neither model text nor the CUDA reference result is an admitted Cohesix job, signed provider outcome or installed-release acceptance. Private redacted observations are retained under out/burn-in/20260926-m28g-60min-mcp-a2a/ on the Mac and /mnt/nvme/cohesix-dev/m28g-burnin-05e/ on the Jetson.

Under m28g-host-clients-as-built-alignment, the NeMo kit now refuses a zero-exit native evaluation when a fixed dataset row has no completed model response in Toolkit’s bounded profiler CSV. The CLI preserves the private trace and reports evaluation_failed with a nonzero process exit; it does not interpret model text as a provider result. The focused NeMo kit and release wheel tests passed 12 cases and scripts/check-generated.sh passed for this source repair. The fresh d3e64749f wheel at SHA-256 252e8e7739ec47c2820b08466008bd40197f623920d3bf08bb81f4bfd5b39ada passed the compiler-selected source verifier and installed on the Jetson with pip check. The refused service request now exits 2 with native_profile_complete=false; the supported local MLX-LM route exits zero with a real LLM_END and the expected answer. These installed-client runs remain direct model evaluation, not a Pi admission or release result. The Pi diagnostic image remains bound to its earlier 05e1facdd source.

Timed physical Pi protocol diagnostic

The Pi 4 GENET Queen at source 05e1facdd008240a12dfe133c033912dcc87d5f5 ran from a RAM-only TFTP image with ID 8e4212e7a8282c0657ad873b97f36a519dc6bf292b16c4554631bdb0680c7251 and SHA-256 3a58570dfc69007c5a002cec7ef22840435581e7a17d4c4482d9bc4b51da64ac. The private selected manifest SHA-256 is bff6b94e8f5bf66e1521360b400fa6bb0f30330a9310598ac0fbeb225c5a6085; master, MCP and A2A are all enabled. Host SHA-256 and Pi RAM CRC checks, the target [BUILD] marker and the first authenticated TCP /proc/boot read bound the live target to those bytes. No SD image was written.

The requested 60-minute read/protocol diagnostic ran from 2026-09-26T08:51:13Z to 09:51:13Z, 3,600.008 seconds. Its 60 minute-cadence cycles, 19 operator health checks, ten Jetson availability samples and six delegated read-ticket renewals all passed; there were zero failed cycles and zero gateway reconnects. Authenticated MCP tools/list returned HTTP 200 with tools=[], and the A2A Agent Card returned HTTP 200 with skills=[] and version 1.2.0-beta. Planned quiet windows occurred at minutes 20–21 and 40–41. The retained event log SHA-256 is 86229d76b0ab17e29f052f4e8516946be45751a2ca182ddf06fa992aefc3a159. The run-owned Mac gateway, protected SSH tunnel and TFTP server were stopped after the finish event; the Pi RAM image and private evidence were preserved.

This is PASS for the executed physical Pi authenticated-read and empty-discovery diagnostic. The requested burn-in and M28g release profile remain INCOMPLETE: the Mac gateway had no co-located standing ledger, so no MCP-only or A2A-only useful job could be admitted and no native task/provider result or recovery was measured. The direct Jetson CUDA/NeMo and Mac Metal checks above have separate host-only proof. The candidate source after the NeMo repair is d3e64749f; the timed Pi image was built from its 05e1facdd predecessor and is not final-source target qualification. Full source-bound staged, installed-client, native package, Worker VM CUDA and release gates remain open. The private out/burn-in/20260926-m28g-60min-mcp-a2a/ directory retains run-record.md, burn-analysis.json, the event log and host observations.

Operator skill portability checkpoint

Under m28g-operator-skill-portability, the nine repository-managed operator skills were checked against Mac and Jetson Ubuntu 24.04 execution contexts. All nine passed skill-creator quick validation; 31 local document links and anchors resolved; 13 Bash snippets passed bash -n on each host. The host integration inventory retained 47 advertised surfaces, 29 dependencies, nine playbooks and six use cases. scripts/check-generated.sh and git diff --check passed. This is instruction and source consistency evidence.

The Mac source-built cohsh completed mock Queen /proc/boot, schedule and lease inspection from /tmp, including an absolute executable path containing spaces. Mac source-built coh help and mock evidence commands worked from /tmp. On Jetson, retained 1.1.0-beta coh created and inspected a mock pack from /tmp; a separate stale development binary correctly refused a compiled policy hash mismatch. The pinned NeMo Agent Toolkit 1.9.0 executable worked by absolute path in a non-login SSH shell where nat was absent from PATH. Mac Codex CLI help exposed MCP stdio and Streamable HTTP setup; that help alone did not establish the two Cohesix HTTP headers or a live job.

The instructions now route copied skills to same-version guides, resolve installed executables without a source-directory assumption, select one gateway process for the Queen connection, and require both private headers and separate verified subjects for multi-user claims. These checks cover two hosts and two observed agent/client families at limited capability scope. They do not prove Claude, Gemini or other vendor clients, cross-user isolation, admitted MCP/A2A work, installed Release B, useful native provider outcomes or release reliability. Those require the remaining installed-client and live acceptance matrix at one final source and package identity.

Full-plan pressure defect and bounded repair

Discovery task: m28g-release-b-qualification. Restoration task: m26e-host-worker-integration, for the Root admission and host-agent delivery boundary found during the selected 26e pressure gate. The initial e1fe34f8c QEMU Test Plan passed Stages 01–04, including authenticated target integration. Stage 05 failed because EX-2026-0024 expired on 2026-09-25; the due-diligence subchecks for audit, advisories, attestation, secret scan and authority floor passed. This exception requires a new human decision and cannot be revived by a source or documentation edit.

The canonical Mac HVF medium-pressure run at that same source accepted admissions but left 56 GPU and 57 LoRA Worker receipts pending beyond its 15-second correlated terminal bound. The host ticket agent logged the first global snapshot gap at expected admission 404 versus observed 405 and then repeated the gap while newer admissions arrived. The 64-line moving Root snapshot can omit a sequence between reads; the agent previously could not recover it, so one lane stopped making progress. The failed run and raw logs remain in ignored out/m28g-pressure-e1fe-r2/ in the disposable pressure checkout. Aggregate HTTP error rate and zero gateway reconnects do not override the failed Worker receipt gate.

The repair adds a Queen-only exact admission read backed by Root’s bounded 256-identity admission window. The agent parses each missing record as a Root-admitted version-2 spec, checks the exact sequence and follows its existing execution journal and lane assignment. Unavailable, malformed, mismatched or 256 or more missing identities fail closed. The gateway sends recovery reads on the control lane without caching the record. Focused tests cover recovery, no provider replay, absent-record refusal, canonical path parsing and gateway routing. This source repair has no pressure, Pi or release acceptance until the selected plan is rerun against its committed identity. coh, cohsh, tools/cohesix-py and benchmark result formats consume the existing ticket identity and result paths; the new read is internal to the agent’s recovery and adds no client action or success state. Their applicable regressions remain in the Test Plan.

The first full Test Plan attempt at repair commit 2202a95a5 passed Stage 01 metadata, generated consistency, formatting, Clippy and workspace check, then failed four cohsh TCP script tests. The authenticated plan environment provided COH_AUTH_TOKEN_REF; the test children set their fixture COHSH_AUTH_TOKEN, which has lower resolver precedence. Each mock server therefore waited for a different AUTH frame and the scripts timed out at attach. This is test isolation, not a target or provider failure. The TCP script harness now removes both higher-priority ambient token variables from each child, preserving production resolver precedence. All six focused TCP script cases passed with the private plan token reference deliberately set in the parent environment. Stage 01 and downstream gates still require a fresh run at the resulting committed source.

The next full-plan attempt at d3100e7cb advanced through the same Stage 01 Rust and target-build actions. Its Python action completed 2,915 tests and 116 subtests, with two failing REST harness argument tests. Those tests cleared the inline console token variables but inherited the plan’s private COH_AUTH_TOKEN_REF, so their expected empty token assertion was false. The tests now also clear that reference in their fixture environment; both focused cases pass with the private reference deliberately present in the parent. Production token precedence is unchanged. The failed Stage 01 attempt is retained under out/m28g/full-plan-d310-qemu/; no later stage or release acceptance follows from it. Fresh full-plan and pressure runs remain required.

The committed 2202a95a5 pressure rerun passed the exact QEMU image, authenticated control operation, fault injections and live Worker preflight, then failed the medium workload: 78 GPU and 55 LoRA operations lacked terminal Worker receipts. The first missing admission was sequence 163. The gateway recorded repeated ERR CAT reason=quota detail=buffer-full for the exact admission read, so the agent could not recover the gap and eventually exceeded Root’s bounded identity window. The read had serialized the admitted JSON as one console line even when it exceeded the line cap. It now uses the existing bounded CAT chunk framing; the production-feature Root test reconstructs a real admitted spec from its frames and passes. The failed medium run remains under out/m28g-pressure-2202/ in its disposable checkout. A new exact-source pressure result is required before this repair can be accepted.

At 0024e6995, the complete QEMU plan passed Stages 01 and 02, including the full Python gate and target-qualified Root build. Stage 03 built both exact images and passed thirteen preceding authenticated scripts, then stopped in shard_1k.coh: WAIT TAIL read a Worker telemetry path before asynchronous Worker construction and received ERR TAIL ... invalid-path. The script’s read-condition contract correctly treats that refusal as final. The fixture now waits for the role-specific WORKER_TASK_READY record on the existing Queen log before reading either sharded Worker path; the public CLI guide documents that order. The six script-catalog tests and generated-consistency check pass. The failed Stage 03 attempt is retained under out/m28g/full-plan-0024-qemu/; it does not establish complete transport acceptance. A fresh source-bound plan run must confirm the changed script.

At 1679e23d7, a fresh canonical pressure run rebuilt the selected Mac HVF image and passed same-boot preflight in separate medium and high boots. The medium report recorded 48,869 operations with zero errors and 256 successful correlated Worker receipts; the high report recorded 72,885 operations with 101 bounded lease-quota refusals (0.1386%, under its frozen 1% budget) and 256 successful correlated Worker receipts. Medium covered 144 GPU and 112 LoRA operations; high covered 132 GPU and 124 LoRA operations. These immutable run summaries are retained under out/m28g-pressure-1679-r2/ in the disposable checkout. The runner then entered the complete staged QEMU plan, where Stage 01 passed metadata, generated consistency, formatting, Clippy and workspace check but failed the shard_1k host-model integration tests. The updated script correctly waited for a target WORKER_TASK_READY record; the synchronous NineDoor host model created its Worker namespace without emitting an equivalent readiness record. The host model now logs a role-specific WORKER_TASK_READY with explicit mode=host-model after creating the namespace. Both focused shard tests pass, including the disabled-alias refusal. The pressure runner did not reach its final collector or establish full Test Plan acceptance at 1679e23d7; the repair requires a new exact-source run. Host CLI, Python and benchmark result schemas are unchanged; the extra mock log record is explicitly labelled and does not satisfy native Worker evidence.

On the Jetson Orin Nano at that same source, a clean native CUDA helper build and bridge build passed all seven direct reference cases. The maintained conformance summary remains INCOMPLETE for production: it has no admitted Worker transport, physical lease/revoke, bridge restart, authoritative result graph, Worker receipt or NVIDIA container-lane proof. The direct result is retained under /mnt/nvme/cohesix-dev/m28g-1679-source/out/m28g-1679-cuda-conformance/.

Earlier exact-source 020a qualification

This earlier checkpoint is clean commit 020a516300b75e55f91606a77106c76ef74f665c, source digest sha256:a098dd349e07e80249019f6b915289baa826dfda865e6c273e26b41d8b117b64. It includes the Root admission-read and host-model readiness repairs above. The selected QEMU and Pi production manifests have the master, MCP and A2A switches enabled; both selected Cohesix and NeMo kit wheels carry 1.2.0b0, the package version for Release B 1.2.0-beta.

The canonical QEMU Test Plan passed Stages 01–04 at this checkpoint, including the full common host gate, exact-image target build, authenticated TCP regressions and REST multiplexer. Stage 05 verified the earlier stage attestations and passed its audit and advisory checks, then failed the release guardrail because EX-2026-0024 expired on 2026-09-25. Its immutable attempt is under out/m28g/full-plan-020a-qemu/. A separate exact-source Mac HVF pressure run passed medium and high: 49,598 operations with zero errors and 72,268 operations with 61 bounded errors respectively, 256 correlated successful Worker receipts in each run and no gateway reconnects. The high error fraction was 0.0844%, within the frozen 1% budget. Its evidence is under out/m28g-pressure-020a-r1/ in the pressure checkout. Pressure success does not renew an expired risk exception. The existing EX-2026-0024 controls also require Jetson KVM medium/high pressure at the selected identity. That lane has not run at 020a516300: the Jetson checkout has KVM access and the pinned QEMU binary but lacks the governed seL4 pressure build and image. Its runner’s read-only preflight stopped on missing out/sel4; a prior-source KVM build cannot be relabelled as this checkpoint’s pressure evidence.

The physical Pi 4 GENET booted a private RAM image with SHA-256 630b35269c483df678c1ef1cad936c04c007297e4587a97fd339e45907fe3ec6 and the same [BUILD] source identity. Its 60-minute MCP/A2A diagnostic ran for 3,600 seconds with four MCP and six A2A original jobs confirmed and acknowledged by the Pi target, native Jetson CUDA outputs independently verified, 122 health samples and no failed health events. A planned publisher epoch handoff at 30 minutes completed with a fresh Worker generation and resource reservation. Three preceding attempts remain recorded as failures: standing-scope attempt exhaustion, publisher epoch removal of the GPU reservation, and delegated caller-ticket operation exhaustion. The fourth attempt’s collector was amended during its timed window to wait for exact Worker READY telemetry and bind the fresh generation; its executed helper hash and the deviation are recorded in the private run directory. The final event log SHA-256 is 49dbd1d5eec70d976e3e1656d3accb2fb92120b2f06fdc99671acd48e033531c. This is a completed diagnostic with a disclosed run deviation, not an unchanged frozen installed-release profile: the Pi ran from RAM, and the run did not cover installed Mac/Linux packages, FUSE, native NeMo protocol clients, PEFT or diverse vendor agents.

The same Pi source/image then passed canonical Stages 01–04. Stage 03 ran 17 authenticated TCP scripts across four fresh RAM boots, and Stage 04 verified same-boot continuity before CLI and Python REST checks through a bound Mac gateway. The first Stage 03 preflight and Stage 01 retry exposed stage environment selection errors; both failed attempts remain retained, and the corrected stage-stable environment produced the recorded passes. Stage 05 failed the same EX-2026-0024 release guardrail. The attempts and exact source/image attestations are under out/m28g/full-plan-020a-pi4/. These RAM-boot checks do not establish a cold SD installed-image series. The run-owned gateway and Jetson services were stopped after their tests.

Direct host checks at this checkpoint passed seven native Jetson CUDA reference cases, an Apple M4 MLX Metal matrix product with independent NumPy parity, and NeMo Toolkit 1.9.0 direct evaluation against the Mac MLX-LM model endpoint. The CUDA reference conformance summary remains INCOMPLETE where it requires admitted Worker transport; the direct MLX and NeMo results also do not establish Cohesix admission or an installed Release B client matrix. Their retained evidence is in the checkpoint’s ignored out/m28g/ roots and the Jetson source checkout.

Under m28g-operator-skill-portability, the operator guidance now distinguishes standing-scope cumulative attempts from caller-ticket operation quotas and requires a fresh Worker reservation after a publisher epoch change. The amended delegation and GPU skills passed skill-creator validation, and the implementation inventory still matches the selected release. The earlier nine-skill Mac/Linux syntax, links and mock checks remain limited to those setups; they do not prove live Cohesix operation across AI-agent vendors. This guidance refinement changes no shipped host tool, Python API or benchmark measurement contract; the 47-surface host integration inventory still agrees with the selected compiler output.

After the failed 020a Stage 05 attempts, Lukas Bower renewed the existing EX-2026-0024 P2 scope on 2026-09-27 through 2026-10-27, as recorded in the exception register. The earlier attempts remain failed. The fresh 522463fa1ade QEMU and Pi Stage 05 governance results above close their staged lane; the selected Jetson KVM pressure control remains unexecuted.

Optional QEMU setup and Python index preparation

Under m28g-macos-native-installer, m28g-ubuntu-arm64-native-installer and m28g-pypi-1.2.0b0-publication, the release setup now makes QEMU an explicit --with-qemu choice. Its Python probe starts the selected four-core HVF or KVM machine briefly, rather than accepting version and accelerator listings alone. The Mac fallback builds from the pinned upstream QEMU 10.1.0 source archive and the selected HVF state-sync patch in a user-owned prefix; it is separate from the native Cohesix installer. Ubuntu ARM64 selects apt’s qemu-system-arm. The repository setup skill and quickstart describe that choice and distinguish a startup check from a booted, authenticated guest. On Mac, installed Homebrew QEMU 11.0.3 advertised HVF but aborted this startup probe; the previously selected patched 10.1.0 passed. On Jetson Ubuntu 24.04, apt QEMU 8.2.2 booted the earlier exact-source 522 guest and completed an authenticated Queen ping. Neither check qualifies every advertised host.

The Python index workflow now selects the future annotated v1.2.0-beta tag on main, builds cohesix 1.2.0b0 from the compiler-selected source inventory, retains wheel/sdist hashes, and hands the exact files to the protected pypi environment for OIDC publishing. The source README carries the matching pip command. A local candidate wheel/sdist build, normalized version check, Twine metadata check and 38 focused setup/release tests passed. An isolated copy of the bundle setup installed the 1.2.0b0 wheel, selected the known-good Mac QEMU binary and passed --check --with-qemu. The new probe also passed on the Jetson’s apt 8.2.2 KVM binary, where the copied setup script passed --check --headless --with-qemu without requiring graphical runtime packages. The first from-source Mac fallback build failed because QEMU’s unused Apple Paravirtualized Graphics module calls APIs obsoleted by the installed macOS 27 SDK; the builder now disables that feature, and its corrected build is pending. Public PyPI upload remains pending the final approved release tag and independently checked public hashes. The generated inventory was refreshed through coh-rtc and scripts/check-generated.sh passed. The current PyPI project still lists 1.1.0b1 and the Cohesix user owner, with no organization owner; the historical organization request and transfer remain pending. GitHub’s pypi environment still requires the named lukeb-aidev reviewer. The new workflow passed actionlint before publication.

On 27 September 2026 the release owner changed Release B from the planned 1.2.0-beta to stable 1.2.0. The source inventory, both wheel metadata, NeMo kit, A2A Agent Card, installer names, release notes, PyPI workflow and build-plan acceptance text now select the stable version. Earlier 1.2.0-beta/1.2.0b0 observations above remain historical evidence for their original source and artifacts. They do not qualify the new 1.2.0 candidate; source-bound release checks, installed package checks and live acceptance must be repeated on the exact stable source.

The release owner also authorized a dedicated Debian signing key. A new passphrase-protected Ed25519 key was created in a separate macOS GPG home, with its passphrase in the login keychain. Its fingerprint is 7E27A4AB355D2EA565718CA41059A516E53B0B70; an independent gpgv probe verified a detached signature against the exported public key. The public key copy is tracked under releases/. This probe establishes key control and verification mechanics, not a signed or qualified .deb package. The encrypted key was imported into the selected Jetson builder’s private GPG home. A Jetson-created detached signature was independently verified on the Mac with gpgv against the exported public key. The native builder can read the passphrase through inherited standard input, keeping it out of arguments, environment values and the package payload.

The pinned Mac QEMU 10.1.0 source builder completed after preserving upstream’s com.apple.security.hypervisor signing entitlement on the staged executable. Its new user-owned prefix passed the four-core HVF startup check and scripts/setup_environment.sh --check --with-qemu; the codesign entitlement was read back from the final binary. Homebrew 11.0.3 remains the observed failing selection on this Mac. This does not yet prove a booted stable 1.2.0 Queen or every supported Mac.

The stable source’s focused Python release, NeMo, installer, setup and archive checks passed after a NeMo wheel filename selector and packaged public-key link were corrected. cohesix-1.2.0 wheel and sdist and the cohesix_nemo_kit-1.2.0 wheel built from the selected inventory; strict Twine metadata passed for the SDK distributions. The selected A2A Agent Card version Rust test and an isolated stable SDK install, import and both CLI help checks passed. actionlint, shellcheck, git diff --check and scripts/check-generated.sh passed. All are source checks; the stable candidate’s installed and live acceptance gates remain outstanding.

M28g remains In Progress. The 522463fa1ade staged results are source and target evidence, not qualification of the installed release. Native signed/notarized Mac and independently signed Ubuntu ARM64 packages, clean installs on every advertised host, installed-client parity, the full live acceptance matrix, frozen adoption and quality budgets, and named release-owner approval remain absent. No exact assembled Release B candidate or publication is claimed. Material AI assistance produced this checkpoint and its scoped validation record.

Pi 4 SD refresh compatibility

Title/ID: m28g-release-b-qualification.

Milestone: 28g / m28g-release-b-qualification.

Goal: Refresh the selected physical Pi card without relaxing whole-disk, partition, filesystem or private-policy checks.

Inputs: Removable 63.9 GB SD card in the Mac’s built-in reader; one MBR FAT32 COHESIX child; exact staged Pi image; the card’s existing private boot policy.

Changes:

• scripts/pi4-image-build.sh accepts macOS diskutil’s observed Windows_FAT_32 content spelling as well as DOS_FAT_32, and still requires the mounted child’s msdos filesystem type, exact parent and label.
• tests/test_pi4_image_build.py exercises both valid spellings through the non-destructive refresh, and confirms a non-FAT32 child or mismatched mounted filesystem is rejected before the card is copied or its policy changes.

Commands: .venv/bin/python -m pytest -q tests/test_pi4_image_build.py; bash -n scripts/pi4-image-build.sh; scripts/check-generated.sh; git diff --check.

Checks: An existing exact FAT32 child is refreshed without repartitioning; other filesystems are refused; the existing private policy is preserved and read back; physical proof remains subject to a fresh exact-image boot and separate hardware acceptance.

Deliverables: Compatible flash validator, regression tests and retained failure logs under ignored out/m28g/. AI assistance identified and repaired the observed macOS spelling mismatch; no physical acceptance is inferred from the focused host tests.

Staged fixture isolation and idle Pi menu

Title/ID: m28g-release-b-qualification.

Milestone: 28g / m28g-release-b-qualification.

Goal: Keep synthetic Test Plan stages isolated from live host selectors and allow the Pi operator helper to read an already idle U-Boot menu.

Inputs: The source-bound eb1b6f03c9d5 Stage 01 attempt, which retained two Python fixture failures when COHESIX_GATEWAY_URL leaked into fake Stage 04; the exact-image Pi first boot, whose U-Boot menu was visible before the serial helper opened the port.

Changes:

• scripts/ci/test_test_plan_evidence.py removes inherited gateway, Pi, QEMU and target-evidence selectors from its synthetic runner environment before applying each test’s explicit overrides.
• scripts/pi4_serial_reboot.py redraws an already displayed choice menu with an invalid, non-persistent selection before reading its state. The helper still refuses to choose a lane until it has read the current menu page.
• tests/test_pi4_serial_reboot.py, the Test Plan fixture test and docs/HARDWARE_BRINGUP.md cover and explain those behavior boundaries.

Commands: COHESIX_GATEWAY_URL=http://127.0.0.1:18080 .venv/bin/python -m pytest -q scripts/ci/test_test_plan_evidence.py -k 'stage_five_refreshes_while_other_stages_resume or upstream_rerun_invalidates_downstream_active_evidence or fixture_isolates_external_target_selectors'; .venv/bin/python -m pytest -q tests/test_pi4_serial_reboot.py -k 'run_returns_nonzero_after_diagnostic_failure or initial_menu_state_reads_current_menu_before_selecting_genet'; git diff --check.

Checks: The three selected staged-fixture tests pass under an inherited live gateway URL, and both selected serial-helper tests pass. The failed Stage 01 attempt remains historical; a new exact-source full Test Plan and Pi boot are still required after this source repair.

Deliverables: Hermetic Test Plan fixture, menu redraw behavior, operator guidance and preserved failed/PASS logs. AI assistance investigated and repaired both observed host workflow defects; no staged or release acceptance was inferred from the focused checks.

Detached Jetson KVM replay checkout

Title/ID: m28g-kvm-pressure-runner-portability.

Milestone: 28g / m28g-kvm-pressure-runner-portability.

Goal: Run the selected non-cleaning Linux KVM replay from an isolated, exact-commit Jetson checkout without presenting that checkout as main.

Inputs: The 1a9dc310db4a guest and clean Jetson source checkout, the canonical --reuse-artifacts --check-only refusal, and the retained m28g-kvm-pressure-runner-portability branch and quiescence contracts.

Changes: The pressure runner applies its main branch rule only to the normal clean-build lane. The replay lane still requires the canonical transferred output paths, KVM/GDB/nm, immutable guest verification, exclusive output ownership, quiescent host and port checks. A focused test exercises detached replay acceptance and preserves clean-lane refusal.

Commands: bash -n scripts/m26e_qemu_pressure.sh; .venv/bin/python -m pytest -q tests/test_m26e_qemu_pressure_cli.py -k 'replay_accepts_detached or quiescent_probe_ignores_runner_ancestors'; .venv/bin/python -m pytest -q tests/test_m26e_qemu_pressure_cli.py (53 passed); scripts/check-generated.sh; git diff --check.

Checks: Both focused tests pass. The initial Jetson check-only attempt correctly refused the detached checkout before this repair. No KVM replay, pressure result or Release B acceptance is inferred from that preflight.

Deliverables: Honest detached replay selection and regression evidence; the exact-source Jetson pressure result remains pending.

A2A catalogue release selection repair

The first canonical 1.2.0 bundle preflight at bc2f7f80399f refused the candidate because the compiler generated a2a_catalogue.json but the release inventory omitted it. This was a real host archive selection defect: the MCP catalogue was selected while the A2A catalogue would not be shipped. The m28g-a2a-catalogue-release-selection repair adds the A2A catalogue to the compiler source inventory and a focused release test. Regeneration changes the inventory, host graph, use-case evidence, MCP/A2A catalogues and provider projections through their normal hash bindings. The earlier bc2f7f80399f QEMU/Pi boots and protocol observations remain evidence for that source, not for the repaired bundle. The focused release suite passed 22 tests and scripts/check-generated.sh passed after regeneration. The compatibility review found no changed CLI, Python call shape, gateway route or benchmark threshold. The Rust and Python provider projections embed the new graph binding, however, so the earlier native host binaries, Python wheel and target artifacts are stale as release inputs. The NeMo kit’s source inventory binding also needs renewed verification. The focused generated-contract and release suites passed 49 tests; cohesix-authority passed 16 unit tests. Exact repaired-source bundles and their installation checks remain outstanding.

Portable production profile for both release builders

Title/ID: m28g-release-b-qualification, discovered while preparing the m28g-ubuntu-arm64-native-installer Jetson artifact.

Milestone: 28g / m28g-release-b-qualification.

Goal: Let Mac and Jetson compile the same exact production manifest and public CAS verification key without recreating a developer-specific filesystem path.

Inputs: The 7c29b8cbb211 QEMU/Pi production profiles; the passing exact source Mac/Pi release preflight; the source-bound JetPack 7.2.1 ARM64 host-tool build; the coh-rtc-authority-profile output containing a Mac absolute key path. The 7c29b8cbb211 physical Pi RAM boot, MCP/A2A smoke and QEMU base transport remain valid only for that original manifest and source.

Changes: The authority-profile CLI requires the deployment public key below the new profile output directory, checks it there, and serializes a path relative to that directory. The private signing key remains outside the profile. The security and authority guides show how to place and move the public key with a generated profile. Focused tests cover a relocated profile and refusal of a key outside the selected output directory.

Commands: cargo test --locked -p coh-rtc --test authority --bin coh-rtc-authority-profile; both QEMU and Pi profile CLI generations into a private ignored output directory; cargo fmt --package coh-rtc -- --check; scripts/check-generated.sh; git diff --check.

Checks: Seven focused Rust tests passed. Both generated profiles retain production authority, MCP and A2A while recording the same relative public key path. Generated consistency passed. New exact-source target artifacts, native packages, installed use cases and release acceptance remain pending; earlier artifact identities are not promoted to the repaired source.

Deliverables: Portable profile generator, documentation and retained focused checks. AI assistance identified and repaired the cross-host build defect.

EX-2026-0025 bounded renewal for Pi release qualification

Title/ID: m28g-release-b-qualification.

Milestone: 28g / m28g-release-b-qualification.

Goal: Restore a current owner decision for the existing Pi direct-GENET P2 unsafe boundary without weakening its controls or promoting an earlier test.

Inputs: The expired EX-2026-0025 and DD-2026-0025 entries; the retained Pi Stage 05 failure at source 7d1ceedb4772; the selected direct-GENET boundary; Lukas Bower’s 28 September direction to fix the boundary or extend the exception through 31 October.

Changes: AI-assisted review found no verified safe replacement for fixed mapped-page atomic access, sealed descriptor reads, suspended-child bootstrap aliases or terminal fault transfer. The named owner renewed the same P2 scope through 2026-10-31 in the exceptions and findings registers. The original unsafe bounds, compensating controls and independent review requirement remain.

Commands: scripts/ci/due_diligence_gate.sh --check-exceptions-register docs/audit/findings.csv docs/audit/EXCEPTIONS.md; scripts/ci/due_diligence_gate.sh --check-blocking-findings docs/audit/findings.csv docs/audit/EXCEPTIONS.md; .venv/bin/python -m pytest -q scripts/ci/test_due_diligence_lifecycle.py; git diff --check.

Checks: Both register gates accept the dated decision, the lifecycle suite retains expired-exception refusal, and the source delta contains no runtime, manifest or acceptance-threshold change. The failed Pi Stage 05 attempt is historical and must be rerun against a clean exact source.

Validation: The register and blocking-findings gates passed; the lifecycle suite passed 36 tests and 51 subtests. The exact-source staged rerun remains pending at this checkpoint. No Pi or release PASS is inferred from the renewal.

Deliverables: A bounded, owner-authorized renewal and separate rerun evidence; no waiver of physical Pi, KVM pressure or installed-release qualification.

Exact Pi stage binding and KVM fault-probe repair

Title/ID: m28g-release-b-qualification.

Milestone: 28g / m28g-release-b-qualification.

Goal: Keep physical Pi proof tied to the image that actually booted and make the pre-READY KVM fault probe reach a target-observable standard fault.

Inputs: The 5de8fdc4ebb6 Pi RAM boot, its passing Stage 03 and Stage 04 records, and the Stage 05 refusal of a missing live runtime/DMA artifact; the Jetson KVM pressure preflight where a GDB zero-x0 marker was followed by a Worker ready-timeout rather than the required Standard fault.

Changes: pi4_gate_proof.sh now accepts the exact --stage-dir used to boot the Pi. Before issuing runtime/DMA proof it checks the staged image digest, stage proof, identity metadata and live [BUILD] marker together. This prevents its former default stage path from silently linking a different image. The external GDB pre-READY probe now redirects the role-bound Worker entry to its existing standard-fault hook. The validator still requires the role- and phase-matched fault from the target; an injection transcript alone does not pass. The Test Plan names the exact stage input.

Commands: bash -n scripts/pi4_gate_proof.sh; .venv/bin/python -m pytest -q tests/test_pi4_gate_proof.py; .venv/bin/python -m pytest -q tests/test_worker_task_evidence.py tests/test_rest_perf_harness.py; scripts/pi4_gate_proof.sh --normalize-only --stage-dir out/m28g/pi4-stage-5de8-dev --manifest out/m28g/release-eval-5de8-dev/pi4-development.toml --log out/m28g/full-plan-5de8-pi4-dev/pi4-stage5-base-boot-and-queenlog.raw.log --require-wired-ready --require-driver-task-proof --runtime-dma-proof-out out/m28g/full-plan-5de8-pi4-dev/pi4-runtime-dma-proof.env.

Checks: All 81 Pi proof tests and 404 Worker/performance evidence tests passed. The exact older Pi boot plus same-boot authenticated Queen log normalizes to PI4_RUNTIME_DMA_PROOF=fresh-pi and COUNTER_PROOF=counter-qualified, with the correct image stage linked. This is a diagnostic run over retained bytes; it is not the active controlled serial/packet capture required by physical performance acceptance. The KVM probe change has focused host coverage but still requires a fresh Jetson run to prove a target Standard fault.

Validation: Source-bound staged tests, physical Pi release profile, KVM pressure and assembled release acceptance must be rerun after this commit. The earlier 5de8 Stage 03/04 and failed Stage 05 records stay immutable at their original source identity.

Deliverables: Exact-stage Pi proof linkage, a fault-probe correction and explicitly bounded regression evidence. AI assistance identified and implemented these repairs; independent target qualification remains open.

Strict-production TCP release input

Title/ID: m28g-production-tcp-release-proof; downstream discovery in m28g-release-b-qualification.

Milestone: 28g / m28g-production-tcp-release-proof.

Goal: Keep production Queen authority strict while producing a native QEMU TCP result that the canonical release factory can bind to the shipped guest.

Inputs: The 4a46989432ae production Pi TFTP/RAM boot and Stage 01–02 PASS; the retained Stage 03 failure at base-telemetry/telemetry_ring.coh, where spawn heartbeat received the production policy’s EPERM for /queen/ctl; the development suite’s separate legacy-control contract.

Changes: run_regression_batch.sh now selects nine production-safe base scripts only under an explicit, validated strict-production mode and gives that boot qemu.production-tcp-smoke identity. The selector refuses an incorrect target, group, action, authority policy, MCP/A2A or standing control. release_inputs.py requires the complete safe script set and matching artifact/action instead of promoting a development Stage 03 result. The catalog, Test Plan and host release guide distinguish the two proof lanes. The full development suite and the production denial remain intact.

Commands: bash -n scripts/cohsh/run_regression_batch.sh; .venv/bin/python -m pytest -q scripts/ci/test_run_regression_batch.py tests/test_release_inputs.py; focused catalog and release-bundle tests; scripts/ci/test_plan_catalog.py validate and check-doc; scripts/check-generated.sh; git diff --check.

Checks: The selector and release-input tests passed 39 cases, focused catalog and bundle tests passed 37 cases with two subtests, and catalog/generated consistency passed. No native strict-production QEMU boot has passed at this checkpoint. The 4a469 production Pi refusal is a failed suite/profile selection, not evidence that production should enable legacy Queen control. All target, installed-package and release acceptance gates remain open until fresh exact-source results pass.

Deliverables: A production-safe proof path and retained failed discovery; native Mac/Jetson replay and separate physical Pi acceptance are pending.

Production Pi NETSTATS response repair

Title/ID: m28g-release-b-qualification; defect discovered during the production Pi 4 authenticated TCP acceptance run.

Milestone: 28g / m28g-release-b-qualification.

Goal: Return the complete bounded netstats response on the physical Pi 4 without changing the existing namespace or log stream batch contract.

Inputs: Source 5cd8e9ef057f and its production Pi TFTP/RAM image at out/m28g/pi4-stage-5cd8-release-prod; the retained authenticated TCP netstats refusal in out/m28g/pi4-prod-5cd8-boot2/netstats.out.log; and the same boot’s successful 67-body-line serial diagnostic in out/m28g/pi4-prod-5cd8-boot2/serial-netstats.log.

Changes: The root event pump previously held synchronous TCP diagnostic bodies in a 64-line log-export batch. A live wired Pi emits 67 netstats body lines, so capture replaced the whole response with a typed bounded-overflow error. Synchronous capture now retains up to the existing 69-line physical-console body bound. Five extra lines are held separately and drained after the ordinary 64-line batch. Log and namespace streams keep their prior 64-line type and limit; overflow still fails with one typed terminal. The host-tool, Python, SwarmUI and benchmark compatibility review found no client grammar, wire terminal, status, schema, or benchmark definition change; fresh Pi replay must establish that the selected diagnostic arrives complete.

Commands: cargo test -p root-task --lib --no-default-features --features release-pi4 bounded_sync_capture_; the same command with release-qemu; cargo fmt --all -- --check; git diff --check.

Checks: Both feature profiles passed the three focused capture tests, including ordered delivery of 67 body lines and one terminal and deterministic refusal at the 70th line. Fresh production Pi image build, authenticated TCP replay, full staged plan and release acceptance remain required for the repaired source. The prior 5cd8 boot and all its valid observations retain their original source and image identity; they do not accept this change.

Deliverables: Bounded synchronous-response repair and host-level regression proof. AI assistance identified the root cause and implemented the repair; independent physical and release evidence remains open.

Linux production host-policy selection

Title/ID: m28g-linux-host-policy-selection; discovery in m28g-production-tcp-release-proof.

Milestone: 28g / m28g-linux-host-policy-selection.

Goal: Bind native Linux host clients to the selected production KVM policy.

Inputs: Clean source 547a65a9fd4a; dedicated KVM guest at out/m28g/kvm-prod-547a-mac; native Linux host-tool manifest at out/m28g/linux-host-tools-547a.json; retained failed Jetson TCP attempt at out/m28g/jetson-production-tcp-547a.

Changes: The first Jetson attempt booted the dedicated KVM guest and passed the authenticated TCP response matrix, then cohsh refused boot_v0.coh: the binary expected policy SHA-256 d5a5d49e62abb9a7353cb8fcb7c8d75df30550174d16546b108516331e8699ad while the selected production artifact supplied f77a645d43347d50eaef1e832b1b729918d1818d1b7f603b4988928036cd321c. The native builder generated host policies from configs/root_task.toml instead of the guest’s selected production manifest. The builder now accepts that manifest as an explicit input, verifies its transfer digest, and records the manifest and generated policy hashes. The release factory forwards its selected manifest and checks native-tool provenance against it. No policy validation or production authority was relaxed. The first selected-manifest rebuild refused before host compilation because the compiler resolved a public CAS key relative to the temporary manifest in the remote build root. Staging the manifest at the source root exposed the tracked fixture public key and production validation correctly refused it. The builder now transfers only the manifest’s referenced public verification key beside the external manifest, independently verifies both digests, and leaves the tracked source tree unchanged.

Commands: bash -n scripts/linux_host_tools_sync.sh scripts/release_bundle.sh; .venv/bin/python -m pytest -q tests/test_linux_host_tools_sync.py tests/test_release_bundle.py; scripts/check-generated.sh; git diff --check.

Checks: Shell syntax passed, 31 focused tests passed and generated consistency passed for the repair working tree. The original KVM attempt remains failed. Fresh clean-source Linux host-tool build, native KVM TCP replay and all assembled-release gates remain pending.

Deliverables: Corrected native builder and factory binding, focused tests, host guide and retained failure. AI assistance traced and repaired the builder mismatch; no target acceptance is inferred from the repair tests.

Exact-image Pi Wi-Fi smoke

At source 547a65a9fd4a, the production Pi image in out/m28g/pi4-stage-547a-release-prod was loaded through TFTP/RAM with three completed transfers and post-reset CRC checks. The SD card’s saved policy remained Ethernet, so a private local Wi-Fi policy was applied for this boot through a redacted serial handoff without writing the card. The retained wifi-smoke-private-ram-attempt3 receipt binds the exact build marker and records a settled Wi-Fi supervisor with no runtime recovery. Same-boot serial netstats reports active Wi-Fi and a bound DHCP lease. The first TCP client attempt was rejected because its transport token was a minted ticket; the retry used the selected transport credential and received OK AUTH, Queen attach and OK PING reply=pong. Retained serial/TCP logs were checked for the private Wi-Fi values and Queen credential. This is a quick same-image Wi-Fi smoke, not the full physical Pi, pressure, repeatability or final SD release gate. The subsequent source change for Linux host-tool selection requires its own exact-source qualification.

Debian publisher key alignment

Title/ID: m28g-debian-publisher-key-alignment; discovery in m28g-ubuntu-arm64-native-installer.

Milestone: 28g / m28g-debian-publisher-key-alignment.

Goal: Ship the public verification key that corresponds to the selected private 1.2.0 Debian signing key.

Inputs: The earlier Ed25519 key probe recorded above, the selected dedicated passphrase-protected RSA publisher key, the tracked public export and install guidance.

Changes: The earlier public key did not correspond to the private key selected for the final native signing flow. Its probe remains a valid historical result for that identity, but it cannot establish trust in a signature made by another key. The tracked public export and reader-facing fingerprint now identify 684071A2AF498008930CF6AEEBB8ECC49397F8C2. The private key and passphrase remain outside the repository. No package or release acceptance is inferred from changing the public trust file.

Commands: gpg --show-keys --with-colons releases/cohesix-debian-publisher-2026.asc; independent gpgv signature verification; focused installer and bundle tests; generated consistency.

Checks: The tracked key parses with the selected fingerprint. A fresh public-only gpgv probe accepted a detached signature from the selected private key and rejected it with the earlier tracked key; its receipt is at out/m28g/debian-publisher-key-alignment/result.json. All 26 focused installer/bundle tests, generated consistency, Test Plan catalog validation and git diff --check passed. Signed-package verification on JetPack 7.2.1 / L4T 39.2.1 and the complete installer lifecycle remain pending at the final source.

Deliverables: Aligned public key, quickstart and release notes. AI assistance identified the trust mismatch; the native package gate remains open.

Linux selected-policy provenance correction

Title/ID: m28g-linux-host-policy-selection; continuation after the native Jetson build at b5d6897e002f.

Milestone: 28g / m28g-linux-host-policy-selection.

Goal: Make the installed Linux policy file and builder provenance identify the policy actually embedded in the native host clients.

Inputs: The completed eight-tool JetPack 7.2.1 native build, its provenance, the selected production manifest and independent native cohsh --check.

Changes: The b5d6897e002f builder did compile cohsh against the selected production policy and the native binary accepted that policy at SHA-256 f77a645d43347d50eaef1e832b1b729918d1818d1b7f603b4988928036cd321c. Its provenance instead reported the untouched tracked default policy at ad39ddfc70f5e0519738403304069384e8fbdd9a9eabc8d3e9ce84edc7d4a769. The compiler had written the selected external TOML under out/ while the builder read configs/generated/. The builder now directs selected cohsh, coh and SwarmUI policy files to their native default paths, and requires the compiled cohsh to accept the selected external policy before packaging. The earlier native build remains a failed provenance attempt.

Commands: bash -n scripts/linux_host_tools_sync.sh; .venv/bin/python -m pytest -q tests/test_linux_host_tools_sync.py; scripts/check-generated.sh; exact-source native Jetson rebuild and KVM TCP replay pending.

Checks: Shell syntax, nine focused tests and generated consistency passed. The updated builder has not yet produced a new native host-tool set or a passing KVM TCP result, so the release gate remains open.

Deliverables: Source and focused regression repair with preserved native failure evidence. AI assistance traced the provenance mismatch and corrected the selected external policy paths; live qualification still controls closure.

Exact 1.2.0 candidate and native installer correction

Title/ID: m28g-release-b-qualification and m28g-macos-native-installer; downstream discoveries in their canonical archive and installer checks.

Milestone: 28g / m28g-release-b-qualification.

Goal: Bind the first exact b68a2c2a1a38 release candidate to selected Mac, Jetson and Pi bytes, and repair defects revealed by qualifying those bytes.

Inputs: Clean source b68a2c2a1a38d6f6b1dfe430845dfc4f0f024298 with source digest sha256:a01c2742283505b0287fffbf6b95dd666fe6088c883e7d55d382b8e3f626c5ba; selected production manifests; independently built Mac and Jetson QEMU guests, native host tools, Python/NeMo wheels and the Pi 4 stage.

Changes: The canonical bundle preflight found serial smoke logs alongside the Pi stage. A separate clean staging copy preserved those logs while packaging the same Pi payload bytes. The Mac and Jetson host qualifiers found that a relative output path failed after the packaged Python smoke changed working directory; release_qualify.py now makes that path absolute. The first Apple installer submission was rejected because seven copied command line Mach-O tools lacked Developer ID signatures. The package builder now requires a separate Developer ID Application identity, signs each copied tool with hardened runtime and secure timestamp before freezing its installed-file manifest, and checks its UUID against the selected archive binary. Installed readback independently requires matching UUID, Developer ID team, runtime, timestamp and exact signed bytes. The rejected Apple submission remains a failed attempt; the corrected package was accepted and stapled. The publication bridge names only the changed release tooling and focused tests as eligible non-runtime changes.

Commands: Exact Mac/Jetson strict-production TCP replays; physical Pi wired and private Wi-Fi TFTP/RAM smokes; canonical release_bundle.sh --check-manifest and Mac/Linux/Pi assembly; release_qualify.py host on both native hosts with pinned QEMU 10.1.0; gpgv against only the tracked Debian publisher key; Developer ID signing, Apple notarization and stapler; focused tests/test_release_bundle.py, tests/test_release_qualify.py and tests/test_macos_installer.py; scripts/check-generated.sh with unfinished release outputs preserved outside releases/.

Checks: The exact Mac and Jetson production TCP results passed. The Pi image SHA-256 5e23f461695a064fea1a0fc9147f54e39dca15c351a0123a7006f65616af962d passed same-image TFTP/RAM boot and quick Wi-Fi checks; the release .img metadata names the same payload hash and its compact MBR/FAT32 layout passed. Both extracted host archives passed manifest, archive, native tools, replay, isolated Python wheel, authenticated TCP and SwarmUI browser checks. The Mac archive SHA-256 is b8fe858b6b3846750397bbd7ceb1e6b8861674b3e5365746ad609bbfcd1dc4ca; Linux is 67a98f0aebc40fde7965420e7a9bfa0ec19a5e5fe1013b6f2db1ce663bba8d3d; Pi is a1cac6542b61521f7be336e83206a3486d78a14aeb4376ea5f496ddf3abcc988. The corrected Mac .pkg SHA-256 is e81886fddec649e1a584695d47744c8d3442719a2e488dd06c399daf71212d0c with accepted Apple submission a4d0164f-8a76-4674-b4f4-ab72e44ad869. The two JetPack .deb hashes are 6f9ed66565100ccc8733495c5217e758897daa758dc043d9867df3e62afc6af3 and ee57cd9372ac8b1579a0d9c6ad92c694d6478f6fdc30ba127da88911bad6f260; their detached manifest signature passed independent gpgv. The focused release tests passed 43/43 and the Mac installer/release qualifier tests passed 30/30. Generated consistency passed after the incomplete release directories were moved to ignored candidate storage.

Deliverables: Exact source-bound 1.2.0 archives, portable Pi .img, signed/notarized Mac and manifest-signed Linux installer candidates, focused repair tests and retained failures. Native privileged installs, installed readback and GUI launch, final SD readback/boot, Jetson pressure, full integrated adoption, owner approval, tag and PyPI publication remain open. AI assistance traced and corrected the path and nested-signing defects; host/package evidence retains its original proof class.

Owner-evaluated installation walkthrough and installed receipts

Title/ID: m28g-installation-and-integrated-adoption; continuation of m28g-release-b-qualification.

Milestone: 28g / m28g-installation-and-integrated-adoption.

Goal: Let the sole developer and tester evaluate the 1.2.0 installation without claiming that his own walkthrough is independent.

Inputs: The named-evaluator adoption schema, frozen budgets in BENCHMARKS, the b68a2c2a1a38 qualified source, and native Mac and Jetson installer receipts for the 869737b03982 publication candidate.

Changes: scripts/ci/provider_m28g_adoption.py accepts Lukas Bower as a person with independent=false and reports evaluation_scope=owner. Other evaluators retain the independence requirement. BUILD_PLAN, TEST_PLAN and BENCHMARKS name the owner exception and continue to require raw walkthrough attachments, three native journeys, GUI launch, budgets and lifecycle checks. The schema advances to cohesix-m28g-adoption-walkthrough/v2 so an older walkthrough cannot silently inherit the changed evaluator rule. The publication bridge classifies only this verifier and its tests as adoption proof tooling, and admits only the exact BENCHMARKS evaluator wording while refusing any changed numeric limit. Host CLI, Python SDK, SwarmUI and benchmark measurement paths are unaffected.

Commands: .venv/bin/python -m pytest -q tests/test_provider_m28g_adoption.py tests/test_worker_task_evidence.py tests/test_release_bundle.py; scripts/check-generated.sh; git diff --check; native scripts/release_qualify.py installer on Mac and Jetson, followed by scripts/release_qualify.py verify with their absolute result paths.

Checks: The 158 focused tests and generated consistency passed. Both native installer receipt verifiers and the strict five-part release verifier passed for the prior 869737b03982 candidate, with exact evidence retained under ignored out/m28g/. Lukas Bower reported that he evaluated the installations and marked them PASS. That report and the installed-byte checks do not supply the still-required raw GUI, three-journey, budget, rollback and uninstall walkthrough. A new publication commit and rebuilt archives will require their own Conditional G comparison before promotion.

Deliverables: Owner-evaluation verifier, regression tests, aligned acceptance text and native installer receipts with explicit proof limits. Material AI assistance traced the stale independence rule and implemented the narrow identity exception; adoption-live and release promotion remain open.

Jetson KVM pressure collector correction and accepted replay

Title/ID: m28g-kvm-pressure-runner-portability; defect discovered under m28g-release-b-qualification and restored in the same pressure gate.

Milestone: 28g / m28g-kvm-pressure-runner-portability.

Goal: Make pre-READY Standard fault injection observable under the selected Jetson KVM debugger without weakening the native fault requirement.

Inputs: Clean target source b68a2c2a1a38d6f6b1dfe430845dfc4f0f024298, selected Linux KVM profile, pinned QEMU 10.1.0 SHA-256 c30027b9805176ad0d85d732c356eac4d7485fab3c1748a797fc193b6bc3a46e, and the failed r2/r3 pre-READY diagnostic attempts under ignored Jetson out/m28g-pressure-b68-* roots.

Changes: The pre-READY collector now redirects the selected Worker PC to unmapped address zero. The previous breakpoint symbol could be consumed by attached GDB under KVM and leave the target at READY timeout. The validator still requires the role-matched target Standard fault and terminal teardown; an injection marker or timeout cannot pass. During-IPC fault and budget injections retain their separate target evidence paths. The release publication bridge classifies only the pressure collector and focused test as a proof-tool delta from the immutable qualified guest source. Host CLI, Python SDK, SwarmUI, provider and benchmark workload code are unaffected.

Commands: bash -n scripts/m26e_qemu_pressure.sh; .venv/bin/python -m pytest -q tests/test_worker_task_evidence.py tests/test_provider_m28g_adoption.py tests/test_release_bundle.py; .venv/bin/python -m pytest -q tests/test_m26e_qemu_pressure_cli.py tests/test_host_integration_run.py -k 'm26e_qemu_pressure or linux_aarch64 or quiescent_probe'; canonical scripts/m26e_qemu_pressure.sh --reuse-artifacts on Jetson with the exact selected artifacts, /usr/bin/gdb, AArch64 nm and the pinned stripped QEMU; scripts/check-generated.sh; local and remote SHA-256 comparison of the final result and both pressure summaries.

Checks: 158 collector/release/adoption tests and 54 targeted pressure/host integration tests passed. The remote pressure collector SHA-256 dd4b415066fd03996235bcd326d1782d07a6543a4c2441adddd88e3f67ca3f27 matched the local repair. The r4 canonical runner exited zero and its final/host-replay-result.json reports PASS with KVM. Medium and high each recorded 256 requested, discovered and READY executable Workers, two minutes of load, strict control errors, no transient retries, a passing 1% error budget and actual error rate zero. The final result SHA-256 is 1513bd3b8f28091fb3067f812cb6f67ed327e969cca929a44e3d2819d56ba6a3; the medium and high summary hashes are 4ced7d7aca597e34e0033fec2f297b4f22ef8bf9b30b312e2f145a8f345a0e2b and 4d774271402e8a479f281ee54d5e3152e27b77d1e476abdb90a6d115a2fc985e. Copies under ignored out/m28g/pressure/r4-evidence/ match the Jetson originals. The test runner stopped its QEMU and gateway processes after validation.

Deliverables: Accepted Jetson KVM pressure record and narrowly corrected collector. This closes the KVM pressure subtask at its original target source, without promoting the still-open integrated adoption or overall Release B. Material AI assistance traced the GDB-consumed breakpoint behavior and verified the real target fault path.

Owner manual adoption attestation

Title/ID: m28g-installation-and-integrated-adoption; owner-directed acceptance-route change under m28g-release-b-qualification.

Milestone: 28g / m28g-installation-and-integrated-adoption.

Goal: Record the sole developer’s personally completed full adoption walkthrough without inventing measurements or confusing it with native proof.

Inputs: Lukas Bower’s 28 September 2026 statements, “I did the walkthrough and marked it PASS,” “I did it manually,” and his confirmation that it covered the full 28g adoption walkthrough; qualified source b68a2c2a1a38d6f6b1dfe430845dfc4f0f024298; the five-part prior-candidate release verifier and native Mac/Jetson installed-package receipts.

Changes: TEST_PLAN and BUILD_PLAN allow a written owner manual PASS for the qualitative walkthrough. The adoption verifier accepts a separate bounded cohesix-m28g-owner-manual-attestation/v1 packet, checks Lukas Bower’s identity, seven attested areas, exact source and both installed package hashes, and emits owner_attested_host. The earlier measured, raw-backed walkthrough route remains available. BENCHMARKS retains every numeric limit but labels the absent manual measurements not-recorded; no numerical result is claimed. The release publication bridge admits only this exact benchmark prose change. Host CLI, Python SDK, SwarmUI, native verifier, target and protocol integration behavior are unaffected.

Commands: scripts/ci/provider_conformance_run.sh --matrix configs/provider_conformance.toml --case m28g-adoption-live --reference-config out/m28g/owner-manual-adoption-reference-20260928.json --host-profile mac-apple-m4-macos27 --state-dir out/m28g/owner-manual-adoption-20260928-final; .venv/bin/python -m pytest -q tests/test_provider_m28g_adoption.py tests/test_release_bundle.py; scripts/check-generated.sh; git diff --check.

Checks: The owner-attested adoption runner passed at the prior 869737b03982 candidate, source b68a2c2a1a38. Its retained private attestation SHA-256 is 2f4cf955b16838388e687add3d8760ebbbb0c835fdbc3d205aa031a0388c1d5c; the final summary SHA-256 is 73a6187068c9f011a3419e6b4f64697a799ff9c4ef084217bc2f0e150b5ca038. The summary explicitly reports measurement_status=not-recorded and native_outcome_reverified=false. The manual statement does not fill the still-open four-mode m28g-integration-live case, native provider results or final publication-byte comparison.

Deliverables: Named owner attestation, bounded verifier and focused regression tests, aligned acceptance documents and retained result with its exact proof scope. AI assistance implemented the separate manual route and verified the package/source bindings; the owner supplied the human evaluation.

Publication-inventory package input correction

Title/ID: m28g-release-b-qualification; package-input mismatch found in final canonical release-manifest preflight.

Milestone: 28g / m28g-release-b-qualification.

Goal: Bind the version-aligned NeMo and Python distributions to the inventory actually included in the publication archive, while retaining their tested source and installed-byte identities.

Inputs: Clean qualified source b68a2c2a1a38, publication inventory SHA-256 5a284e05493cb426fc0133adf043401563884deeab1c73c6c562a29657e87851, the prior installed 1.2.0 package receipts, and the failed preflight using the qualified-source NeMo distribution report. Historical release rows changed the publication inventory digest without changing the selected NeMo wheel.

Changes: The attempted qualified-source verification shortcut was withdrawn: the canonical factory continues to verify the package report against the publication inventory it will ship. A publication-source NeMo report was rebuilt from that inventory; its wheel SHA-256 remains 526a16c54dac26d5a0099ad9ce78a819bb1acedac90d731d89d3214d9f5af4f6. The earlier installed Python wheel is selected for final assembly only after its unpacked members are compared with a fresh publication-source build; its source/package report must remain exact. No host, target or provider runtime code changed.

Commands: build_nemo_kit.py --out and build_python_package.py --out against the publication inventory; unpacked wheel member comparison; canonical scripts/release_bundle.sh --check-manifest --linux --linux-use-accepted-tools --qualified-source-root <clean-b68-source> with the exact retained Mac/Linux QEMU artifacts and passing TCP results, clean Pi stage, selected production source manifest, installed package reports and publication-inventory wheel reports; git diff --check.

Checks: The initial preflight refused NeMo distribution differs from selected release source because its input report named the qualified inventory. A temporary factory change made that preflight pass but produced a bundle whose NeMo report did not match the bundled inventory; that candidate is retained outside releases/ and is not qualified. The attempted code change was reversed. The publication-source NeMo wheel has the same bytes as the qualified wheel. The freshly built publication Python wheel has identical unpacked members to the earlier installed wheel; its ZIP bytes differ, so the installed wheel is retained as the exact distribution input. Final canonical preflight with the publication-inventory NeMo report, installed Python wheel and both installed package reports passed: the compiler inventory, source selection and release manifest agreed, the NeMo verifier reported PASS, and the factory reported Native artifacts and exact compiler-selected release inputs verified (build-only=0). Archive qualification remains pending at this record.

Deliverables: Corrected package-input selection with preserved failed candidate and no factory behavior change. AI assistance identified and reversed the inappropriate verifier shortcut before publication.

Owner-selected 1.2.0 release evidence

Title/ID: m28g-release-b-qualification, m28g-host-clients-as-built-alignment, m28g-installation-and-integrated-adoption.

Milestone: 28g / stable 1.2.0 release acceptance.

Goal: Apply the owner’s 28 September 2026 focused release selection to the exact tested archives and installed packages.

Inputs: Qualified runtime source b68a2c2a1a38d6f6b1dfe430845dfc4f0f024298, the 869737b03982a87e4a14c85a605659b48cbf16cc publication archives, and the existing Stage 05, hardware, host, native installer, pressure, component-live, protocol-control and adoption records.

Changes: BUILD_PLAN and TEST_PLAN select the completed evidence at its original identity for 1.2.0. The three sealed archives and their matching signed installers were copied byte-for-byte into releases/. The release notes describe the selected user workflows and JetPack reference in plain English.

Commands and results:

Distributed archive SHA-256 values are Mac 663ef55d9360aab536dbbee2486770171ea28b89093c01cfaffbf86eb3514d6d, JetPack Linux 28115fa5153938ee64925142706c3eed01b74f9a63e2f5aed3714989801a6674, and Pi 4 128917dbf6f67188f5143c22c0563bfdb82626db894b5ac4f257fc687961e505. The Mac .pkg SHA-256 is 416004f05ff419cca8116255824d2dec8f1c0eed842289c6cace7c4b35ccb77c. The controller and SwarmUI .deb SHA-256 values are 76a76cde3805d0efb4b50aa886d05c0dd917967dd62626edf070acf8d83b1826 and 1f5b5b4dd21bed2e92465d70392e06ac119266a1ba07b1d983b98902ca65c10a.

Checks: scripts/check-generated.sh passed with the release files staged. Regeneration added ten extracted release README rows to the implementation inventory and updated only its dependent graph and provider digest bindings; the existing inventory rows and other graph fields were unchanged. The 27 focused adoption/bundle tests passed. git diff --cached --check passed for the changed build plan and status text; the sealed archive members retain their qualified bytes. The publication bridge passed at clean documentation commit 8596cd315, recorded in out/m28g/publication-8596.json. The current source changed release prose and proof collectors after the qualified runtime commit; selected runtime binaries, image, packages and signed distribution bytes remain the exact qualified set.

Deliverables: Exact sealed release files, the selected Test Plan gate, owner walkthrough attestation and hash-bound acceptance index. PyPI publication and its public distribution hashes receive a separate post-tag record.

Cohesix 1.2.0 publication

Title/ID: m28g-release-b-publication.

Milestone: 28g / stable 1.2.0 release publication.

Goal: Publish the owner-approved sealed release and exact tagged Python SDK.

Inputs: Lukas Bower’s approval of the six artifact hashes and local commit e3fa7657773236e3e7b0147fee9c400771bfad39; the accepted results and distribution hashes above; his separate approval of the CI-built wheel and sdist hashes below.

Changes: Fast-forwarded origin/main to the approved commit, pushed annotated v1.2.0 with peeled target e3fa7657773236e3e7b0147fee9c400771bfad39, and published the GitHub Release with the six distribution files, installer manifest, signature and publisher key. GitHub reports the six asset SHA-256 values listed above. The release is neither a draft nor a prerelease.

Commands and results: GitHub Actions run 36417223728 built and checked the tagged cohesix==1.2.0 SDK, including Twine and an isolated installation smoke. The retained distribution artifact’s SHA256SUMS, source commit, package metadata and wheel RECORD were checked locally. After Lukas Bower approved these exact files, the reviewer-gated PyPI deployment succeeded on attempt 2. Public PyPI downloads were read back and independently hashed:

Checks: Remote main and the tag’s peeled target matched the approved commit; GitHub asset digests matched the local sealed files; the pypi environment reviewer gate accepted the owner’s exact-file approval; both public PyPI file bytes matched the reviewed hashes. The PyPI wheel is a tagged-source build and has its own bytes, distinct from the wheel inside the native archives.

Validation: GitHub main, tag and asset identities matched; GitHub Actions completed its tagged build and reviewer-gated upload; both public PyPI file hashes matched the approved CI distributions.

Deliverables: Cohesix 1.2.0 and cohesix 1.2.0 on PyPI.