Wire exact native CPU capability and disposable OpenCode trial

[erikleblansch]

Latest actual trial: invalid tool-call JSON, no completed coding loop

Run 37683661957/job 113005829633, attempt 1 on Code 45510055bfe4f8ead010e403e701e5c710efe499 / core 9e9086ee873f7162f498e8647e6f6f438273e6ce, failed. Build, conversion and provisioning complete. The task ends after 475,006 ms within the unchanged execution deadline. Core records one result_observed and two conversation_native_json_syntax diagnostics: repeated validation of one report, not two executions. One request returns one cleanup-confirmed incomplete result (invalid_output / invalid_model_output). Deadline, token-limit and wire-truncation counters are zero.

No usable native turn, tool call, approved read/edit/test or command completes. The fixture is unchanged and independent verification is not reached. Core cgroup peak is 1,535,148,032 bytes within 11 GiB, with zero OOM kills and swap; this is not a complete inference-worker peak. Reported runtime/project, guest-unit/private-data, QEMU/scratch and owned CI-host cleanup pass. Normalized route/DNS observations match; raw byte hashes differ, so neither byte-identical host state nor a full firewall audit is claimed. Raw model output is not exported.

Source review confirms the new mandatory verification scope is checked before result_observed. The syntax rejection is inside the recognized tool-call body. Exact malformed bytes and a unique generation/template cause remain unknown. Native generation is greedy without grammar constraints; this is a source fact, not a proven unique cause. Parser/permissions, model/prompt, resource limits and deadlines are unchanged. No repeat model trial is authorized by this result.

Original artifact 11511695654: 9,982 bytes, ZIP SHA-256 93edd2b2b92db78a807a1de11a431afb38ede10443b965c9c4fe523eac75573f. Original job log: 68,937 bytes, SHA-256 e0b7f8f2c8073ade7ee168c2ef9a4efddc2f5b710f9686c8621fbdf68eb2cd17. Original archives/logs and all consumed one-shot receipts remain unchanged.

Exact-source checks pass: Code source-contracts 37680333983, core Quality 37680238438 and core CodeQL 37680230981. The ordinary topology job was skipped, not runtime proof. Both PRs remain draft.

Retained earlier checkpoints

The wording below records older checkpoints, not current CI or trial status; the latest actual outcome is above.

Current: explicit native verification scope, source checks passed

Code 45510055bfe4f8ead010e403e701e5c710efe499, tree 491a92e3cc793a88e2946ffce3a1801af1e3e4b5, pins exact core 9e9086ee873f7162f498e8647e6f6f438273e6ce. Code's only runtime selection change is that core pin; documentation preserves the actual failure below.

The native worker carries its measured initial conversion-source provenance instead of reading unused source shards twice. Execution GGUF still has full before/after hashes, and Rust requires the exact new verification scope. Source integrity, converter tensor checks, model, prompt, resources, governor, deadlines and ordinary PyTorch checks remain unchanged.

Root repeated all 36 native fixture tests with the exact retained core/loader source and zero skips. Source-contracts 37680333983/job 112994401813 passes on this exact Code head. Core's fresh exact-source CI remains pending. No model has run on this correction; no completed coding loop is claimed.

Latest actual trial: verification deadline, no completed native turn

Run 37671030306/job 112962546447, attempt 1 on Code 2b2687fe / core 36d616fb, failed at compute_deadline with cleanup-confirmed execution_budget_exceeded. Build, conversion and provisioning passed. Internal generation recorded 44 tokens at 518,788ms, followed by verification at 519,120ms; no verified result reached OpenCode before the 603,331ms task termination. Last worker state was paused under quiet_hold, with all 31 controls acknowledged. No unique overrun cause or total pause duration is inferred.

Zero native turns, tool calls, approved reads/edits/tests and commands completed. The fixture remained unchanged and independent verification was not reached. Core cgroup peak 9,466,974,208 bytes stayed within 11GiB, with zero OOM kill/swap. Reported runtime/project, units/private-data, QEMU/scratch and CI-host cleanup passed. Normalized route/DNS observations matched; raw hashes differed, with no full firewall-audit claim. Raw model output was not exported.

Original artifact 11505764017 ZIP SHA256 935e115a86abf1da83fe0e3d68ebac90d37ecbe399c7dcc79f055ff26dc33e83; original raw 68,786-byte log SHA256 369407954fff82728d2ff41d5d0e229c78726ea8ae487bf2112964138ca492d8. Original failures and consumed one-shot receipts stay unchanged. A new runtime requires separately reviewed exact-source preflight; no rerun has been dispatched.

Historical source and trial record

The entries below retain their contemporaneous descriptions; “current”, “latest”, “pending” and “no trial” refer to those older checkpoints. The newest exact-source status is above.

Current integrated diagnostic candidate — source CI pending

Candidate 2b2687fec5376b6e49f8266dff71cc7935845e99, tree 3588e69290343a49dd28334c60c829031ee6d133, parent 2c667d450afb354ffd287938ff4092bd59a83972, pins exact core 36d616fb6d79a36fe7c8ee95c4ad84bd56b561ae. The core integrates main's verified public-filter milestone and preserves both conflicting documentation sections. Native inference and diagnostic bytes are identical to bccd; the only Code runtime selection change is the exact core pin.

Root repeated all 89 focused OpenCode Python checks on the diagnostic follow-up and 36 native checks after the final merge pin, with exact retained sources and zero skips. The core still rejects invalid output and adds only closed syntax/data/eof/io diagnostics, never parser messages, field names, positions or raw output. Model, prompt, resources, deadlines, Torch defaults and permissions are unchanged. Exact-source CI is pending. No new model execution or successful coding result is claimed; previous failures and consumed dispatch receipts remain preserved.

Latest actual native trial: failed on invalid conversation output

Run37654339995/job112905525075, attempt1 on Code8b5d5b6ad2d764a503cffb508d5c176711c36840 with coref4cd8bd7ca310794dfc73cfcd096d6baedcf73bf, completed FAILED. Exact source/runtime/native build and full BF16 model provisioning succeeded. The guest reached the OpenCode task and retained result_observed plus two conversation_native_json diagnostics. The provider submitted one request and returned one incomplete result (invalid_output / invalid_model_output), with zero completed turns, tool calls, permissions, edits or tests. No successful coding or full model-turn result is claimed; the underlying output-shape cause remains under source review. Raw model output was not exported.

Core cgroup peak was4,555,878,400 bytes within11GiB, with zero observed OOM kill or swap. Reported owned units/processes/private-data/QEMU/scratch cleanup passed. Normalized route/DNS observations match, while raw IPv6-route bytes differ; byte-identical parent snapshots or a full firewall audit are not claimed. Task-memory and graceful-core-stop fields are absent, not passing or zero observations. Original artifact11499080343 ZIP9982B SHA256 c7d3ed9150e8eb8e7d58a65ed46b0179c997a74f2191c621cf0109552278b6dd; original decoded job log SHA256 fd62b12408eee3e4bbf6c994a84a9cede4fd088806103525782cf428f1ff6610. Both and the consumed once-only helper/receipts are retained unchanged. No automatic repeat dispatch or main merge. Original model, prompt, resources, deadlines and acceptance gates remain unchanged.

Previous source: bounded native loader compatibility

Candidate 8b5d5b6ad2d764a503cffb508d5c176711c36840, tree ee8d94e1bc5d55072fbfae1efb421053daed08af, parent d48801a66f69b094b7f41d99240ee9d7418eb91c, pins core f4cd8bd7ca310794dfc73cfcd096d6baedcf73bf. The core serializes complete tensor validation and services owner controls between tensors, without changing the model, prompt, precision, resources or deadlines. Single-tensor latency is not claimed to be bounded by the owner-control interval.

The guest consumer now loads the exact source-pinned sibling contract afresh, restores import state even after errors, and delegates build-manifest v2 provenance and actual compiled-loader binding checks to the core. The patch is included in wrapper-source hashes; the copied provisioned build must retain its exact digest and pass the same v2 checks. Old v1 build manifests are rejected; the library ABI remains version 1. Torch defaults and its core pins remain unchanged.

Root independently repeated all 89 focused OpenCode tests with explicitly retained exact sources, with no skips; an independent review repeated all 36 native tests and found no blocker. Two new retained-source tests skip in standalone CI when those explicit inputs are absent, never downloading them. This is inert compatibility evidence, not successful model execution. Exact Code source-contracts37650492562/job112892239423, core Quality37648810363/job112886443122 and core CodeQL37648795706 pass; the separately authorized trial above failed. All historical failures and consumed receipts below remain retained.

Previous actual trial: model loading failed

The one-shot trial37641148934/job112860150266, attempt1 on Coded48801a6/core773d6377, completed FAILED. Build/conversion/provisioning pass with the same full BF16 model. All20 owner controls are acknowledged and preparation reaches model_load/begin, then reports worker_other. The retained diagnostic does not identify the precise native failure; no model turn, tool call or coding task completed. Six failed runtime requests have confirmed cleanup. Core cgroup peak8,106,295,296 bytes is within11GiB, with zero swap/OOM kill.

Reported owned units/processes/private-data/QEMU/scratch cleanup passes. Normalized route/DNS observations match; raw IPv6 route bytes differ in the explicitly excluded counter column, so byte-identical snapshots are not claimed. No raw model output was exported. Original artifact11492792961 ZIP SHA256b2621c8918f1304848fcc4a7b9e736fab2379cdf19c29de9ab11606e5c9e9cfc; original job log SHA2564e32650af9e373e5e28ec93663777f1fe833feee419a5fe604fcfdfa6b7ec438. All prior failures and consumed helpers/receipts are retained; no repeat dispatch or main merge.

Source review separately found unbounded asynchronous tensor-validation launches in the pinned loader despite the two-thread inference limit. The source-only candidates above implement bounded, complete validation and closed failure diagnostics; full-model runtime remains unproved. This is not a retrospective proof of the trial's cause, and no resource limit or validation requirement is being relaxed.

Scope and current source

Previous candidate d48801a66f69b094b7f41d99240ee9d7418eb91c, tree cf627fb2e8873d89342ef7f251af207a37b9e0e2, stacked on feature/terminal-compute-budget at 714904a3ec452b8e10d8b40be6ac85107606ca3b. This Code counterpart to core PR212 pins exact core 773d6377e5263c24a9263f7b97b5c9ae739b3f74.

Prior8a70 Code source-contracts 37630557310/job 112823395397, prior10fd core Quality 37627671053/job 112813538681 and CodeQL 37627663225 passed. New d488/773d source checks passed. Root repeated 50 OpenCode fixture checks and 22 smoke checks after the pin update. These source checks do not prove native inference.

The latest follow-up reads only a versioned, bounded closed conversion-failure diagnostic. Missing, malformed, duplicate or unknown diagnostics remain explicit and cannot produce success; raw stderr/paths/private context are not exported. Root independently repeated 61 targeted Python fixture tests after the final core pin. One new diagnostic trial37623495917/job112799334236, attempt1, was accepted at 2026-10-07T12:46:57Z after fresh all-green source/workflow/no-duplicate checks and independent review. The diagnostic trial completed FAILED at model provisioning. The converter exited with status 0; verification rejected its output (stage=verify, failure=contract). The closed receipt does not identify the first failed assertion. No model was admitted and no inference/coding task ran. All reported guest/process/private-data/QEMU/scratch cleanup gates passed, with scoped routes/DNS unchanged. Original ZIP SHA256 50ad3ba19fe9811e132f2dad7ee3c49407d5fcecc86a65a816d7a0637eab996e; original job112799334236 log SHA256 24bcefeb3279b24e011008fecca2037a0808dac2af82587527c1d378759c86a3. Original once-only receipts remain retained; no repeat dispatch.

The conversation parser requires the complete negotiated native backend/greedy-policy pair for qwen3-4b-instruct-2507-v1. Partial or mismatched capability fields remain refused. Existing Torch defaults, frontend permissions and final private summary are unchanged. Core retains executable/artifact authorization; this is not frontend attestation or an application peer scheduler.

Original results preserved

The original Torch trial37228603308 remains failed: no completed first forward and no coding result. The earlier coref7e2c3b9 Quality run37235233676/job111532896805 failed on four full CLI-tree checks. Core7308371b fixes the duplicate Clap group with a unique group ID and regressions; it does not establish resolution of the original model stall.

The first explicit native trial 37239050033/job111543869498, on Code77fc08a8/core7308371b, also remains failed. It stopped at native-source-build, before provisioning or a coding task. Its active/exited helper observation had no retained resource counters or manifest. Guest units/private files and outer QEMU/scratch cleanup passed; scoped routes/DNS comparison remained unchanged.

• Original ZIP SHA256: ed5f8c699806fc1f4f7438ca2ffc3ded2e26a8cd7d95e445204cd9c815ac2e07.
• Original log SHA256: 717fe94bb4b1ae05c281b7f4096e9dc4cbc05268b4f933461bd39d8cfc8e041e.

The original exception/substate was not retained. An inert reproduction and systemd source support empty-cgroup pruning as a possible cause, not a uniquely established retrospective diagnosis. The consumed original dispatch receipt/helper is preserved and will not be reused.

Follow-up and unchanged trial boundaries

The native helper now joins the compiler and verifies provenance, then stays alive for a bounded parent ready/release handshake. Kernel SO_PEERCRED binds both ends to the expected process/user/group. Parent observes the exact live helper as the sole remaining cgroup member, revalidates the manifest and requires actual resource counters before release. Successful exit still requires status0 and an empty unit. Source fetch, build and handshake share the original absolute1800-second build deadline; missing, mismatched or late evidence fails closed.

• Guest-only exact llama.cpp source 7fe450e19305b828c199d602c23a8337aaa1f03b; no prebuilt binary or host source build.
• Same original 4B weights/full prompt/JS task, approvals, greedy policy, two threads and600-second request budget. No quantization, smaller model, trimming, fallback or inferred answer.
• Same12GiB guest,13GiB no-swap QEMU,14GiB host admission,20GiB/1800-second provisioning,7200-second VM and6600-second SSH limits.
• Native build remains two jobs/nice19/11GiB/no-swap/1800seconds. Actual swap/OOM observations must remain zero. All owned units/process groups join before cleanup.
• Closed substate/stage/exception categories reveal no raw logs or private paths.

Root independently repeated all57 targeted Python fixture tests; a separate reviewer repeated the27 native checks and reviewed all four changed files. Tests include actual cross-process sockets/credentials, failed exits, deadlines, stale/mismatched messages/manifests and resource refusals. The unchanged consumer/session code retains its earlier47 passing Node tests.

Remaining evidence

The original follow-up trial37614132814/job112768290137, attempt1, is FAILED at model-provision/native_conversion, process status1. The source build and ready/release handshake now complete successfully with a verified manifest/library; native-build peak is 1,031,565,312 bytes, with zero observed swap/OOM. Those counters apply to the build, not the later conversion. Wheel-graph and runtime-import checks passed before conversion; the catch-all failure marker does not identify the failed conversion substage or underlying exception. At that earlier revision model provisioning, inference and coding-task execution remained unproved; the newer result below now demonstrates provisioning only. All reported units/process-group/private-data/QEMU/scratch cleanup gates pass, with scoped routes/DNS unchanged. Original ZIP SHA256 d4b32d66e90d4c76568b2fa989cbab5e49bbb736283868a03a81f9ae0303a254; original log SHA256 6e561f75a15560a56f3d98c1a49e518c52ad543646ea1463c7f475e5ec2d0ce6. Both are preserved alongside the original pending/accepted dispatch receipts. No rerun, second dispatch or main merge. The bounded closed-stage diagnostic follow-up is now implemented and source-tested as described above, without changing model, converter pins, resources or acceptance gates. It is not a demonstrated conversion repair.

Previous metadata correction trial: failed

The one exact-source trial 37631532611/job 112826781182, attempt 1, on Code 8a7034d5 / core 10fd16dc completed FAILED during the OpenCode task. Source build, conversion and model provisioning now pass: the admitted original BF16 GGUF is 8,051,283,488 bytes, SHA256 d7934f6fa06f9781bc15540f0e2b83f89b1895e86440a4e97af4f0434311d1bd.

The worker reaches verify_files/begin, then fails compute_control_ack_deadline: capacity requests an I/O-pressure pause (19.12); two controls were issued and only the initial resume was acknowledged. Source review identifies missing execution-thread control checkpoints in the full original-weight hash loop, also present in subsequent native bundle/post-execution verification. The new core pin adds execution-thread checkpoints between unchanged 1 MiB hash reads, including initial, native bundle and post-execution verification. All 138 targeted core worker/native checks pass, independently repeated; complete integrity verification, pressure thresholds and deadlines are unchanged. A single blocking read can still exceed the deadline. This is source/control-path evidence, not another actual model trial.

No native model turn, tool call or coding task completed. Six runtime requests failed with cleanup confirmed; those are OpenCode behavior within this one trial, not repeated workflow dispatches. Core peak is 2,303,844,352 bytes, with zero OOM kill/swap. All reported owned-unit/process/private-data/QEMU/scratch cleanup gates passed; routes/resolv.conf snapshots match and no raw model output was exported. This is conversion/provisioning progress, not inference, private peer execution or full coding proof.

Original artifact 11487417682 ZIP SHA256 78c4732774fc93dae9872df0460e2addf7f3849a6a190f803069fbc7d6a6e9c8; original job log SHA256 ef7d5c7bbeeebf497ec9a42881746a782b36f7c7bddc9febef11d7af373cb44f. All earlier failures and consumed one-shot receipts remain preserved. No repeat dispatch or main merge.

[erikleblansch]

5c6248b adds explicit owner-level outputConstraint=qwen_tagged_tool_json_v1 selection through runtime configuration, isolated session argv, provider connection negotiation, submitted conversation and exact result binding. No tools means no constraint field; absent option preserves old frames/argv. Unsupported cores fail before runtime startup, without fallback. The model may still answer normally and all existing tool approvals remain required.

76 targeted Node tests and10 inert Python session tests pass (real HTTP/Unix framing, synthetic core/native replies only). Independent review checked compatibility with core PR212's manifest/request/IPC contract. The new source-check workflow includes the Python selector tests. Actual trial pins/provisioning are not yet changed: native source compilation/sanitizers and the joined source checks must pass before the real OpenCode scenario is rerun. No successful tool/action/coding claim is made.

[erikleblansch]

Commit689c7dd wires the explicit native selector all the way into the real OpenCode trial runtime and pins core9d044921f7e684737f9b10868429de42bc5416c9. New native trials require build3/backend2 capability provenance and selector-bound task receipt2 at both guest and outer verification; original Torch argv, receipt1, prompts, permissions and budgets remain unchanged. 83 targeted checks pass (38 native Python with the exact retained core parser,24 smoke Python,8 CI compatibility,13 JS); one separate retained-loader check is explicitly skipped. No local build or model execution.

Core source-only normal/native constraint checks pass; the wider unchanged upstream sanitizer test still records its known quantized-kernel failure. Core Quality38672934132 and the Code source checks gate any new real inference trial. This is a wired candidate, not proof that the earlier invalid JSON/model task failure is resolved.

[erikleblansch]

Original actual trial 38675278393 (Code 6aa6291 / core 9d044921f7e684737f9b10868429de42bc5416c9) is audited and still FAILED, with meaningful partial progress:

• Native build, conversion and provisioning pass with backend v2/build v3 and qwen_tagged_tool_json_v1.
• Seven submitted requests yield one completed function call and six execution_failed errors. The real OpenCode runtime completes one read action; no edit, test, complete native turn or coding acceptance follows.
• No invalid-output or execution-budget error is reported. The only retained core event is result_observed; the six failures' causes and full request ordering are not established.
• Runtime/project, provisioning-group, guest/private-data, VM/scratch and owned CI-host cleanup complete. Scoped normalized route/DNS observations match, not raw IPv6 route bytes.
• All 33 bundled Code sources, the core archive and six native wrappers match the exact pins. No raw model output exported.

Artifact 11946278747 ZIP SHA-256: 487bee78bb4857c6a5b48df7d69df21ceeb608ded9d9264122668480f62afa74. Original log ZIP: 5ff8a4055bf966f192f80b366cbeb8428d77793c6ee6525b3109eecf292af36e. README/evidence updated at 70c72fe. No retry dispatched; the next narrow change closes the pre-supervisor/post-report diagnostic gap without changing the model, prompt, budgets, retry policy or acceptance.

[erikleblansch]

The closed task-boundary diagnostic is now verified on exact core 5607f1177dc230a71414e7468c221b950f43bca6: Quality 38678859785 passes all 3 new diagnostic Rust cases and 13 output-constraint cases (workspace: 3561 distinct passes, 7 ignored). Namespace skips remain skips. Original log ZIP SHA-256: 60e43c4a0a63a250b4af601245d9e41772135149e47663a730a511866347d1b9.

Code e39267343612ed8b594d15db3e161465c25b8c2f pins that source. Code CI 38681259931 passes 405 checks with 4 explicit skips; all 114 source blobs match the merge checkout. The extra skip is the synthetic namespace test on a runner lacking bwrap, not actual sandbox acceptance. Original log ZIP SHA-256: 0b9d42a5f0526ada26cf56166b2698d48b83510a3d6ffb5d5950a113d3b9e93c.

Exactly one actual native follow-up 38681471469 is now running against that Code/core pair with the unchanged explicit Qwen4B profile, native backend, prompt/output/resource/time limits, task and acceptance rules. It is not a success claim or a presumed fix for the six execution failures. Core ledger-only commit 11bffbb6 records this evidence; the tested execution pin remains 5607f117.

[erikleblansch]

Published candidate 7f451e6: only the native core pin and its test expectation change to 3630ff7905f401f26245f2e6cf4f0f01d7950549. Models, ABI, Torch pins, memory admission, deadlines, privacy and acceptance guards are unchanged. Targeted inert consumer checks: 81 unique passes and one explicitly skipped retained patched-loader-source check; no native/model/VM execution. Automatic source checks: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38694623103 (pending). This is limited authorization to prepare one isolated development VM experiment after source checks, not core Quality acceptance, merge or release approval. Original core Quality 38688104130 remains failed on the storage uniform-lifecycle lock test; its cause remains unresolved. No new actual experiment has been dispatched.

[erikleblansch]

Original source-check audit: run 38694623103 succeeded for candidate 7f451e6. Tested merge checkout 2c64f01 has tree 6510052b89e4bcb97a97ae0bf4d766e906548690; all 114 tracked file blobs/modes exactly match the candidate. Original results: Node 296 pass/1 skip; offline Python 6 pass/1 skip; OpenCode Python 93 pass/2 skips; constraint session 10 pass; syntax passed. Total 405 pass/4 explicit skips (retained upstream prompt, bubblewrap, retained exact core source, retained patched-loader source). Log ZIP: 60958 bytes, 12 entries, CRC verified, SHA256 3d21b00011ca752ba081d900e5eda70aa31214c9ae6ca3fb6898cca9b02a252d. These are source/protocol checks, not actual inference or core Quality acceptance. Original core Quality 38688104130 storage-lock failure remains unresolved. No new actual experiment was dispatched by this publication/audit.

[erikleblansch]

One bounded development VM experiment is now running: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38695131367 . Exact Code 7f451e6, core 3630ff7905f401f26245f2e6cf4f0f01d7950549, model qwen3-4b-instruct-2507-v1, backend llama_cpp_bf16_v1. This is an isolated development experiment after verified consumer source checks; it is not merge/release approval or core Quality acceptance. Original core Quality 38688104130 remains failed on the unresolved storage uniform-lifecycle lock case. No retry, fallback, model/resource change or additional run is authorized by this checkpoint. Original runtime receipts and cleanup will be audited when this same run completes.

[erikleblansch]

Original actual run 38695131367 completed FAILURE (attempt 1), exact Code 7f451e6 / core 3630ff7905f401f26245f2e6cf4f0f01d7950549. Closed evidence shows one completed model-generated function call/read, followed by six execution_failed responses with compute_model_memory_headroom. No edit or test completed; fixture_changed=false; no actual coding acceptance. No invalid-output/token-limit/wire-truncation diagnostic and no observed OOM kill/swap. This proves memory admission still refused subsequent requests, not which memory source limited admission or whether the advisory cache hint ran/reclaimed anything. Provider cleanup confirmed for all seven requests; guest units/provision joined, private data removed and guest host state unchanged; QEMU joined, VM scratch removed and original host route/DNS observations unchanged; original ACL/user-manager and build cleanup succeeded. Source checkout and 33 Code input files, six native wrapper files, full archived core source, runtime/build/manifest/backend bindings were verified. Artifact 11955096952: 10192 bytes, 11 JSON records, SHA256 f194f9d173436c95818ea2bcdca2327c35bea8a2d3896c65329c563508e1d5c1. Original log ZIP: 40078 bytes, 19 entries, SHA256 2e4eccae4819c9656c71f482a1278b88ebb120813e94bcf004c6f7d47a88fcd5; both ZIP CRCs valid. This remains a failed bounded development experiment, not core Quality acceptance or a lock-fix claim. No retry dispatched.

[erikleblansch]

Checkpoint: bounded native memory diagnostics candidate published as 8b214da19ca61bfad631a5dcc09415bfeb7940c1, pinned to core 0091e75f49552ab5c5a0934cab6ae08463c45b17.

The original admission decision and GGUF advice policy are unchanged. This adds opt-in closed measurements, not a demonstrated memory fix. Native guest receipts may use the existing 256 KiB export ceiling; Torch retains its original 64 KiB path and golden contract. Export also rejects pretty-print expansion beyond that ceiling.

Local inert checks: 75 passed, 2 explicit skips (11 CI + 27 smoke + 37 native); no local Rust, model or VM execution. The maximum-count synthetic v2 receipt was accepted by the real collector and exercised through bounded receipt loading/export; it is not runtime evidence.

Core Quality 38702928794 has passed formatting, native Python, and the targeted Rust native/cache/admission/sampler gate; full Quality is still pending. Code source checks 38703873328 have started for this candidate. Both complete gates remain required before a new actual model trial. No trial was dispatched, and this is not merge/release acceptance. Earlier actual admission refusals and the earlier failed Quality record remain historical evidence.

[erikleblansch]

Original source-CI audit complete: run 38703873328, attempt 1, SUCCESS for candidate 8b214da19ca61bfad631a5dcc09415bfeb7940c1.

The original checkout log identifies PR merge 2ae8835aeeb40f9867900dced8754aa2ddd409b9 (parents 714904a3ec452b8e10d8b40be6ac85107606ca3b and the candidate), tree ca8d6c3b9d2b7ae04f6d2d3088d000d7be7fdddb. All 114 file modes/blob identities match the candidate exactly, including the five changed files.

Original results: 410 passed, 4 explicit skips, zero failures — Node 296/1, offline build-contract Python 6/1, OpenCode Python 98/2, explicit constraint-session Python 10/0. Syntax also passed. This workflow installs/runs no model, native runtime or VM; the skipped optional environment-dependent integrations are not inferred as working.

Original log ZIP: 61,525 bytes, 12 entries, CRC valid, SHA-256 e0fb481b4ae0a8e9d8dc917cf10bc54850da96c2073cfe96af97409d643d4a2a; this workflow publishes no artifact ZIP. Run/jobs/artifact metadata and original logs were retained for audit.

The Code source gate is satisfied. Full core Quality for exact 0091e75f49552ab5c5a0934cab6ae08463c45b17 is still a separate required gate before any new actual model trial. No trial, merge or release was initiated by this audit.

[erikleblansch]

Native core pin is now 60a7386bd5dedf614a08be497b41963f842c9512 in Code 0c28e81. This changes only the two pin lines from 8b214da. Original Source CI 38705689397 passes 410 tests with four explicit skips; checkout 5ce2714 has the exact candidate tree 107de91abda939256e70e604cf4b7a9de4908e3d (all 114 file blobs/modes match). Original log ZIP SHA-256: aa8cf3ffa10e0da81a4ded8a1ca19d7c379c41e8711510c963ee7fd1eaae895e.

The previous core Quality stopped on two Clippy findings after targeted tests passed. Their narrow corrections are being validated by core Quality 38705500842. No new model trial has started: full core Quality must also pass, and model/task/resource limits remain unchanged.

[erikleblansch]

Started exactly one actual native inference run: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38709341304 on Code 0c28e81, pinning core 60a7386bd5dedf614a08be497b41963f842c9512. Dispatch was gated on successful exact core Quality 38705500842 / CodeQL 38705495404 and matching Code Source CI 38705689397; no older inference run was active.

Core original logs and checkout tree were audited before dispatch: VOLPAROSSA/volparossa#212 (comment) . The Qwen3-4B BF16 profile, native backend, two CPU / 12 GiB guest / 11 GiB core limits, 10.5 GiB admission floor and task remain unchanged. This is a live candidate, not a successful edit/test result; the earlier failure 38695131367 remains part of the evidence.

[erikleblansch]

Original evidence audit for actual trial 38709341304, exact Code 0c28e818c71c0b66b180e303d36439760e992596 / core 60a7386bd5dedf614a08be497b41963f842c9512:

• Run completed failure, not a live observation timeout. Artifact 11961451056 SHA-256 364b548f95fa0ca9db0d65757698ec67f8b581fb6e000eecc6e5783cc44f654f matches the API; original log ZIP SHA-256 77c788115712e0f41e155b7ab02c874fe67141c21e84b65972cefad86fc39d33. Both ZIPs pass CRC. Checkout, all 33 Code input blobs, six native-wrapper blobs and the core archive bind to the exact sources; their original validators accept the closed records.
• The actual runtime completes one model-generated read. Six later requests report compute_model_memory_headroom; no edit, test or fixture change occurs.
• The first deciding admission has 11,619,966,976 bytes of host headroom versus the unchanged 11,274,289,152-byte requirement. All six subsequent refusals identify the service's own cgroup: deciding headroom is 9,844,723,712–9,844,887,552 bytes, not unavailable measurements or an ancestor limit.
• The single bound model-cache advisory reports success after confirmed worker cleanup. Its immediate after-sample is distinct from the subsequent admission samples; roughly 1.94 GB remains charged as file cache. This identifies the limiting input, not which files account for those pages or whether they are safely reclaimable. No OOM kill occurs.
• Runtime, project, VM and owned-runner cleanup are reported complete; VM evidence is scoped to the recorded route/DNS comparison and does not claim raw-byte host equivalence. No confidential peer-execution proof is claimed.

Model, task, CPU/memory limits and admission guard were unchanged. Next: inspect the source-weight/GGUF cache lifecycle using the retained evidence before another actual run; no repeat has been started.

[erikleblansch]

Native source-cache candidate is now pinned by Code 4fc3692 to core 2dd38983a504d2b3a8ee0e5805e95a44234c621d. Only the exact native-core pin and its assertion changed; model/runtime sources, task, output constraint, memory-admission floor and CPU/VM budgets are unchanged. Targeted local contracts: 38 pass, one explicit skip, including inert checks against the exact pinned core. Hosted Code source checks 38722953808 pass 410 checks with four explicit skips. Core Quality 38722825812 and CodeQL 38722771032 remain pending; no new model trial is started until those source gates are audited. The last actual outcome remains one read followed by six admission refusals (38709341304), not a completed edit/test task.

[erikleblansch]

Original audit of actual trial 38726624377: FAILED / coding task incomplete. This was the single trial after audited core Quality 38722825812 and CodeQL 38722771032 passed; no replacement trial was started by this audit.

Source binding: original checkout logs identify Code 4fc36928bb12ae16e609e45d39e7a6d66b0b0272 and core 2dd38983a504d2b3a8ee0e5805e95a44234c621d. The clean Code source receipt binds tree 3b974bed459995f794e68db8e56475d1edd3eddc. The Qwen3-4B BF16/native model, task, output constraint and resource limits were unchanged.

Observed progression, kept distinct:

• Seven requests, one completed model result/function call and one completed read observed by the tool diagnostic. No edit, test, fixture change or complete coding-task acceptance.
• First execution: admitted with 11,647,827,968 bytes headroom; a result was observed. The existing GGUF cache advisory then reported success after successful supervisor cleanup.
• Second execution: also admitted, with 11,616,268,288 bytes headroom against the unchanged 11,274,289,152-byte floor. It subsequently stopped with compute_memory_pressure, not an admission rejection. Its closed state records decision=cancel, constraint=memory, 440,078,336 bytes headroom, below the unchanged 512 MiB runtime guard. The last observed substage is tokenizer_load/complete; no later prompt/model-load/generation substage is retained. The terminal state does not retain which memory scope supplied that minimum, so this is not attributed specifically to self/ancestor/host at cancellation. OOM-kill count is zero.
• The remaining five requests were refused at admission with compute_model_memory_headroom. These deciding samples explicitly identify the service cgroup, with 834,162,688–834,437,120 bytes headroom. About 10.93 GB is charged as file cache, but aggregate counters do not identify the files or establish reclaimability. The second execution's GGUF-hint record is supervisor_error, with no advisory/after-snapshot. The separate source-weight-hint marker was not enabled/exported in this historical collector; its individual calls cannot be inferred from these receipts.

Original closed diagnostics validate: 8 events, 1 execution state, 9 memory records; no truncation or unrecognized records. Runtime/project cleanup, joined provision group, empty guest units, removed private guest data, joined QEMU and removed scratch are all confirmed. Recorded route/DNS comparisons are unchanged; the receipt also reports raw route/DNS equality and no firewall modification. No raw model output was exported, and no confidential peer execution or successful whole-task claim is made.

Original artifact 11969982126: 11 JSON files; SHA-256 98fd3597191fe21cf13574b83cac3892d7717aaefe2accf26a4a2475b0a1429e, matching the GitHub artifact digest. Original log ZIP SHA-256 026460c95f4238a5db2f650f126ea38787fc1e1a5331d101a6c78fc3fb2a5568. Both ZIP CRC checks pass; original run/job/artifact metadata is retained.

Next is a source-bounded review of verification/cache lifetime before GGUF preparation. The observed failure does not justify lowering admission/runtime guards, extending deadlines or claiming cache ownership from aggregate measurements.

[erikleblansch]

Follow-up candidate 3a5c1f5 changes only the two native core-pin lines to core3327115e07d2c4fa976b2d0c625729a68df0aaf6. The original failed actual38726624377 and its closed audit above remain unchanged.

Core332 adds a native-only verified-source descriptor lifecycle before GGUF preparation; all complete source/GGUF verification, memory floors, execution bounds, model/runtime pins and tool/workspace permissions remain intact. It is a candidate for the observed preparation-time pressure, not proof that specific pages were reclaimed or that coding now completes.

Core targeted checks:191 inert casesPASS, independent reviewPASS. Code targeted native suite:37PASS,2explicit retained-source skips; the exact new-core import case was then separately run andPASS (the retained patched-loader case was not rerun). No native library/model was loaded locally. Automatic source-check run38733022967 is terminalSUCCESS; its original source-bound audit follows. CoreQuality38733078778 and CodeQL38732695543 are the only new pending core source gates. No new actual has been dispatched and no budget was raised.

[erikleblansch]

All exact-source prerequisites for Code3a5c1f59b282952e378e9525f2cf47da43ee19b2/core3327115e07d2c4fa976b2d0c625729a68df0aaf6 have now passed and their original logs/checkout bindings were audited: Code Source38733022967 (410PASS/4SKIP), core Quality38733078778 and four-language CodeQL38732695543. Core Quality original ZIP SHA-256 dbe843ee87550cf3b38fed5b839e27ef395f684b358b1f0ec625818015b684ac. Exactly one actual read/edit/test follow-up is now running: https://github.com/VOLPAROSSA/volparossa-code/actions/runs/38737284583 (job116429700011, created2026-10-15T12:46:59Z). Same Qwen3-4B-Instruct-2507 profile, llama_cpp_bf16_v1, 180-minute job/115-minute actual timeout, task and resource/admission limits; no fallback or weaker guard. Both candidates remain frozen. Previous38726624377 remains a failed two-admission/one-read trial with no edit/test. This source-cache lifecycle candidate and its new pending trial do not establish confidential peer execution or general coding quality.

[erikleblansch]

Original hosted trial 38737284583 is terminal FAILURE on Code 3a5c1f5 and core 3327115e07d2c4fa976b2d0c625729a68df0aaf6.

The pinned source/build/preflight stages passed. During guest-inference the owned QEMU unit reported oom-kill / signal 9; SSH returned 255 and no guest-result.json was recovered. This does not establish which memory boundary caused the kill, which guest phase had been reached, or whether the new source-cache hint ran. It is not the same observed guest memory-guard failure as the preceding trial.

QEMU join, removal of the private VM scratch/overlay, unchanged observed host routes/DNS, runner restoration and build cleanup are confirmed. Guest-internal worker join/private cleanup are not independently evidenced without the missing guest result. No successful edit/test/model-execution or confidential-execution claim.

Original artifact 11974713652 contains 10 closed JSON records, SHA256 031678c85a91bc8210c5dff121b8a1037cdb31e9519bc5a19cfdd36ce39d0cf2 (API digest matches); log ZIP SHA256 358264ace79ac02eee6d355a8d095a0245b20d5ed14623a5bad14af48c480d39. Both ZIP CRC checks pass. Exact checkouts, 33 Code source blobs and the core Git archive were independently audited.

No duplicate actual is running. Next scoped work is bounded host-cgroup diagnostics in a separate Code candidate, without changing model, task, resource limits, cache settings or privacy.