fix(code): negotiate terminal model budgets without automatic retries

[erikleblansch]

Scope and source

PR #5 is merged into main. This candidate normally merges that milestone, preserving its banner and original public-proof evidence. Exact Code head 714904a3ec452b8e10d8b40be6ac85107606ca3b, tree 457494284a3301acd5c177e3cd9ac2c3c19d684f. Both native smoke profiles pin reviewed core 6a517b576baa17e7329661ee0476d1848081d114 (core PR #210); distinct model identities and resource budgets remain unchanged. Public-code core 2a1b4ad3 and immutable driver f27576eb are unchanged.

• Native preflight and every provider request negotiate execution_error_version:1.
• Only a negotiated, correlated, admitted conversation can retain a cleanup-confirmed execution_budget_exceeded; map it to HTTP422 without inventing a model result. Busy/other failures/uncertain cleanup retain prior behavior. Legacy clients remain compatible.
• Fix a separately reproduced native configuration bug: pinned OpenCode drops agent.temperature:0 unless the model advertises temperature capability. Both custom profiles now advertise it, so the existing zero request actually selects greedy_v1. No prompt, model, budget or permission change; this is a generation-policy forwarding correction, not a solved 4B stall.
• Correct the synthetic error fixture to echo only the policy actually present in its request, never invent one from negotiated support.

Actual native-runtime checks

Existing source-built pinned OpenCode v1.18.34 was run through the production launcher/provider with synthetic core/model replies and no model/download/host-network changes. invalid_output, wire_truncated and execution_budget_exceeded each issued exactly one coding request, no retry and no approvals; all observed greedy_v1. Session/process/socket/temp cleanup passed, original disposable project unchanged. The budget failure has one confirmed cleanup and no completed/incomplete model result.

The first probe failed because the synthetic fixture falsely injected a policy; its result remains failed. After the fixture correction, an intermediate probe passed but observed no selected policy. After the separate config fix, the final local report native-errors-budget-greedy-df481-20261004.json passes with SHA256 a08f26ab2c61b2653949dadf06f45b9402066383ee651d573a5b14f9623e3786. It was captured on the explicitly modified df481 candidate, not represented as an unchanged earlier commit or as real inference.

76 focused Node checks pass (one optional upstream-source test skipped because that worktree lacks the upstream checkout; root independently read the exact pinned upstream implementation), plus 22 smoke contracts. The earlier 93 focused consumer checks remain scoped evidence. Historical default-setting fixture bytes are unchanged; only the explicit temperature-capability delta is allowed by its regression. Core post-merge diagnostic/subscriber tests also pass.

No 4B coding-quality, native-editor or confidential private-peer success claim. Original native run 37218791500 remains failed. The new exact-source trial below also failed; its improved terminal handling is distinct from coding success.

Source-quality follow-up

The original core a5246942 Quality run 37225595367 remains failed: strict Clippy rejected two 103-line diagnostic functions. Its original job 111504476222 log is preserved with SHA-256 1ef3aca56b4052b9b56e2f9a469543e07cfdb1718d5a8706750c286f40f6d41c. No model trial was dispatched for that pair.

Core 6a517b576baa17e7329661ee0476d1848081d114, tree 8a2b328bc35d1455d9f88d87671881b05a3757aa, extracts the existing snapshot parser and moves byte-identical Python test assertions into a constant. It does not change validation, model behavior, resources or acceptance criteria. Eight targeted diagnostic tests and formatter/diff checks pass locally; no local modern-Clippy pass is claimed. The Code follow-up changes only two native source pins, their two test expectations and these documented bindings; its 22 smoke contracts and all five source checks pass. All five required core source checks now pass, including strict Clippy, workspace tests, namespace-proof reporting and the integration harness. The explicitly manual KVM source-workflow job was skipped as designed; the informational aggregate CodeQL check is neutral while all four actual analyses pass.

Original native 4B result: failed, with a more precise boundary

After all exact source checks and one guarded HTTP204 dispatch, run 37228603308, job 111513342870, attempt 1 completed FAILED on Code 714904a3ec452b8e10d8b40be6ac85107606ca3b / core 6a517b576baa17e7329661ee0476d1848081d114. Original ZIP SHA-256 fabcc8ba08f47f28f85614b645679654b70f83f02c85c4f8be8722c0d1d5978c (9,154 bytes, 11 files); original job-log SHA-256 6de3c4d41e1346582dc045248b25bbbf353125691f7c4e2ec191be8a8bb0e27f (67,048 bytes).

• Actual pinned runtime build and all 50 provisioning downloads completed (8,312,769,415 bytes; original model revision cdbee75f17c01a7cc42f958dc650907174af0554).
• One real provider request reached generation with a 4,572-token prompt. The first forward began at 19,672 ms; the retained terminal observation has no completed first forward, no first token and zero generated tokens. The 600-second compute deadline ended the request. CPU observation reports AVX2 and no AVX512-BF16/AMX features; this is not proof of the exact slow kernel or root cause.
• The genuine runtime received exactly one execution_budget_exceeded, with one confirmed cleanup and no retry. This proves the scoped terminal-budget behavior against a real core/model attempt, not successful inference. There were no tool calls, read/edit/test approvals, completed native turn, edited fixture or passing independent check.
• Runtime/project, private inputs, provision/service units, QEMU and scratch cleanup passed. Disposable-host owned changes were restored; normalized route/DNS observations match while raw route/DNS bytes differ. No raw model text was exported.
• The fixture's original_test_unchanged:false is its untouched pre-verification default: this run never reached the independent-check phase. The legacy pack flag code_contains_uncommitted_changes:true is not evidence of an unexpected checkout edit; the separate source receipt binds the exact clean Code checkout.

The original 372187 failure remains unchanged. This new run was not restarted; no follow-up dispatch is authorized by this result. Native coding quality, native-editor success and confidential private-peer execution remain unproven.