Conan low-cost observer boundary, 2026-11-06

The lower-cost observer remains unimplemented and unqualified. The authorized separate benchmark boat was created, but its first SSH-key registration returned HTTP 403. That operation was not retried, delegated, or bypassed through another command route. No benchmark, emulator build, generated guest control, original Conan session, production repair, or gameplay observation ran in this task.

The shipped first failure remains exactly sequence 71332, endpoint 1/device 136454160/generation 1. Writer 114805, thread 1, EIP 0x41ccee, return 0x41cd09, and source 0x074ff2d0 identify the already-singular source at upload. Source/copied/lowered/executed bits agree. sourceOrigin remains null; construction and lifetime are unknown. This provides no CPU/timer diagnosis. Final-copy and native-matching x87 work is already proved and was not repeated. Guest evidence, matrices, timing and fatal guards were unchanged. Native player movement and a distinct second response remain unproved.

Evidence is retained in scratch/runs/20261106-conan-low-cost-origin/. Read-only review covered CLAUDE.md, the Conan TODOS entry, both required notes, the previous observer helpers and retained control evidence. static-inventory.json records nine evidence hashes and the retained store-site inventory. The previous observer rejection remains 10.255131x: 2479.562500 ms versus 241.787500 ms median, exceeding the unchanged 1.20 bound.

Static review found 2391 wrapped stores, including 68 in $uop_fast, 266 in the uop compiler fragment and 25 in register helpers. Each wrapper can enter a resident-owner scan with linear global getters, owner/table translations and overlap checks. Other possible costs are per-dispatch/uop metadata, decoder host records, forced native service markers and broad per-store snapshots. The old identity helper also leaves CDP debugging enabled; effects on tiering were not measured. These are cost hypotheses, not measured attribution.

The inherited native-entry marker deliberately invalidates selected-writer history. Its qualification script's nativeEntries == 0 assertion cannot prove positive selected-writer native coverage. Selected hot-path stores and complete destructor entry/return history remain missing requirements. Removing that invalidation or omitting a route would not qualify a replacement.

Small run-local preparation comprises the cost profiler draft, diagnostic ablation draft, provider helper and five copied observer-helper inputs totaling 58452 bytes. No source tree or media archive was made. The ablation variants explicitly disallow qualification/original launch. Only Node syntax checks of the two new JavaScript drafts and Python AST parsing of the provider helper passed; none was exercised as an observer. No performance result, native-code capture, architectural passivity result or new coverage result exists. No locked build was needed for this documentation-only change.

The coordinator grant began 21:13:54.531655 UTC and ended 21:58:54.531655, including cleanup. Repeated disk-floor checks stopped provisioning before any API call until coordinator lossless compression restored headroom. One fresh idempotent create then returned bx_z9jjgb59, created 21:17:25.751, original expiry 22:02:25.765 UTC. The SSH-key POST failed once with HTTP 403. The helper retained status and route but did not read the error response body; its exact reason is unknown and was not recovered by repeating the operation.

Under explicitly delegated cleanup, normal stop completed at 21:17:59.255 UTC; provider state is archived, stop status completed, forced false, current expiry null. No SSH tunnel, Chrome process, server or profile was started. The local runtime holder released its locks after this receipt. Final cleanup, source SHA, relevant binary diff, new-file/helper patches and publication metadata are in the run; result.json is published last.

Recovery begins by reviewing boat-api-failure.json, boat-sshkey-attempt.json, boat-created.json, boat-state.json and boat-evidence-retrieved.json. The security-flagged operation remains held: do not replay it, reenroll the key through commands, resume the archived boat, create a replacement to evade the refusal, or delegate that action. A provider/owner resolution and fresh explicit coordinator authorization are required before benchmark work can resume.

After that resolution, pin the current Git/source/module closure and reconstruct the small helpers from retained patches. Profile the rejected observer and measure diagnostic ablations before selecting a replacement. Profiling cells must stay out of timing medians. Same-source BOTH-module generated full-shell native controls must prove computed fixed work, first MOV/FST/FSTP pre/post, owner epochs, selected hot-path and destructor coverage, architectural passivity, all existing negative guards, and cost at most 1.20x. Keep 16 lifetimes/2048 calls/8192 details and strict source/owner/generation/module correlation. Missing coverage must reject qualification explicitly. Only full PASS followed by a separate explicit coordinator original grant can authorize at most one materially changed original session. No such PASS or original grant exists here.