Ignis 0.8 one-GPU preflight evidence

This directory is the sanitized public projection for the 2026-10-26 focused functional preflight and supported tool-choice follow-up. Protected native artifacts remain authoritative; the files here retain both phase outcomes, public-safe identity, projection provenance, failure analysis, and explicit gaps.

Campaign boundary

• Campaign ID: 2026-10-26-ignis-080-preflight
• Capability: direct text/tool functional preflight and mandatory-tool request-policy qualification
• Executable preflight revision: dc576b25083d7cfdab1554db89379feee82346b5
• Phase-two harness patch: 0c9b211fd6bf8d49493a2db46f5def4ebfe9a887d2470d8a74f6f37331b46206, reviewed at c727ff01f34d2d7564fab63d94203fec68890baf
• Evidence label: functional
• Outcome: phase-one tool_choice: "auto" expectation unmet at 3/4; the same runtime/model/recipe later passed the separately frozen tool_choice: "required" normal-thinking path across 3/3 mechanism cells, 5/5 full checks with 9/9 observations, and a 4/4 seed-18 repeat
• Decision label: no-promotion
• Promotion boundary: this packet does not authorize lifecycle, route, client, or promotion changes

The candidate used Ignis 0.8.0 with the pinned Qwen3.8 Flash Next artifact on one RTX PRO 6000 Blackwell Max-Q. The second equal card remained protected and had no model allocation in retained candidate snapshots. Phase two changed the harness/request policy only; no engine, image, model, profile, or recipe changed. This is not a performance, capacity, SWE, coding-quality, or engine-comparison campaign.

Common artifacts

• artifact-manifest.json - finalized ten-role ledger and exact public-byte identities
• source-registry.json - dated official and local source provenance
• summary.json - bounded machine-readable outcome
• friction-log.md - failure, count correction, warnings, and deferred fix-forward work
• restoration.json - starting/ending state and independent post-run checks

Workload, identity, and raw evidence

• plan.json - corrected nine-request/six-phase execution shape, frozen controls, and zero-retry boundary
• configuration.json - public-safe model/runtime/recipe, hardware, and resource identity
• preflight-native.json - sanitized projection of the native preflight/v2 artifact; synthetic reasoning excerpts and tool outputs are retained
• engine-request-summary.json - public-safe admitted/done event correlation for all nine requests
• resource-summary.json - bounded GPU, host, cgroup, and OOM observations
• restoration-smoke.json - sanitized native routed smoke projection after the phase-one incumbent reload and readmission
• restoration-smoke-phase-two.json - sanitized native routed smoke projection after phase-two exact restoration
• independent-restoration-review.json - public-safe independent Sol acceptance of operational restoration only
• tool-failure-analysis.md - request construction, tool_choice semantics, validator boundary, and engine correlation
• tool-usage-contract.md - pinned Ignis 0.8 source contract for auto, required, thinking, and the request-policy boundary
• tool-choice-followup.json - all three mechanism cells, selected full preflight, independent-seed repeat, exact request accounting, transport hashes, and interpretation limits
• prepared-followup.json - separate 12-cell coding plan status; zero cells were submitted
• projection-ledger.json - protected-source hashes, transformations, public paths, and disclosure boundaries

Result boundary

The phase-one shared-prefix tools check missed its frozen expectation at 3/4. The failed response was ordinary assistant prose with finish_reason: "stop" and no tool_calls. Since the request used tool_choice: "auto", the response is not itself a protocol violation, but it fails the harness's explicit instruction and all-four tool expectation. No root cause, engine defect, or model defect is assigned.

The original private plan and pre-live review counted six HTTP requests. Six was the number of sequential phases; the actual command made nine HTTP requests because the tools phase submitted four requests and the tool-result phase submitted two. Native observations and engine completions agree at nine.

Phase two retained that result and tested the documented mandatory-tool path. All three seed-17 mechanism cells passed. The frozen selection rule chose required with default thinking; its full run passed 5/5 checks and 9/9 observations, and the seed-18 tools repeat passed 4/4. Engine terminal IDs 1-16 formed a complete set. Literal HTTP 200 was retained for 12 captured tool requests, not all 16 observations. The passing one-request seeded auto cell does not rescore the older unseeded 3/4 batch or establish auto reliability.

The prepared coding follow-up remained unrun because its explicit evaluation route and protected-source prerequisites were not available through a supported path. It remains a separate zero-cell plan and contributes no model measurement.

Publication state

Phase-one operational restoration is complete: exact incumbent recipe/image/ revision/GPU-order identity, candidate absence, byte-identical before/restored router custody objects, managed readmission, one passing routed smoke, idle engine state, and zero OOM delta were retained. The first smoke command was rejected locally for a missing explicit auth environment variable and made zero requests; the r2 command used the supported credential resolver and made the one passing request. Router configuration-file bytes were not inspected and no router configuration write was attempted.

Phase two also passed exact operational restoration: all compared incumbent identity fields matched, the candidate was absent, router custody was unchanged, managed readmission and one routed smoke passed, the engine was idle, and container OOM counters stayed at zero. The single smoke is not full quality or strict-JSON acceptance, and mounted router configuration bytes were not independently inspected. Independent exact-head publication review is the remaining publication gate. No compact social publication was requested, so the final publication-summary role is not-applicable with that reason.