Wave 1 friction and dispositions

Consolidated during execution; original failures and reviews remain authoritative.

Observation	Disposition and durable follow-up
Baseline manual GPU snapshots used the wrong bind address	Original failures retained. Continuous Prometheus history recovered; future snapshots use verified gateway address. Record collector identity in run configuration.
First router quiesce lacked the scoped credential	No mutation occurred. Retried with managed token resolution in memory; value never retained. Keep managed secret resolution in the operator workflow.
Authenticated Flare model identity and trace capture were not integrated	Added protected file-backed key reader/mount checks, observer model-probe credential support, fail-closed trace CLI and regression tests. Source snapshots and independent reviews retained. Native fleet hook release remains open.
First live verification immediately after each swap saw stale/missing scrape state	Retained failed attempts; current local engine/model/generation checks passed after discovery/scrape catch-up. Treat model registration and fresh successful scrapes as separate gates.
Remote engine intermittently unavailable during instrumented admission	Current local checks pass; full-fleet verifier was blocked by the separately owned remote endpoint; final restored-state checks later passed20/20. No remote lifecycle change. Record local acceptance separately.
Incumbent C4 request145 exhausted2048tokens in reasoning with no visible answer	Original failure retained; independent review allowed remaining frozen diagnostics without retry/replacement/budget change. Entire C4 arm and overall ranking are ineligible. Extended sweeps deferred.
Prior reports called retained_ns a frozen submitted estimate	analysis-contract-v2.md corrects this to retained shape-cost observation at selection. Earlier reports remain historical; final analysis uses v2.
Cache byte units were initially mislabeled	cache-units.json and reviewer erratum retain the correction:9374826496bytes=8.73099GiB=9.37483GB. GPU snapshot eviction counts are not host eviction counts.
Primary output length and API sampling are uncontrolled	Diagnostic stack workload only. Prompt seeds do not become API seeds. Separate1024-token fixture validates length/timestamps but has no semantic oracle. No formal comparison/qualification claim.
Common coverage/friction files were not copied before first request	Execution plan existed; common ledgers were consolidated later. Preserve that chronology and do not retroactively claim the formal frozen artifact-set protocol.
New profiling may exceed planned outage reserve	Complete bounded core blocks, then stop expansion by the original reserve. Record omitted lanes explicitly and restore exact incumbent.
Restoration hash check initially compared recipe-monitoring.json to the pre-integration snapshot	Failure retained in restoration-protected-initial.json. Only the authorized protected Flare key-file reference differed; current bytes match authenticated-recipe-monitoring.json. Original router/topology and six discovery hashes pass. Future restoration checks distinguish approved persistent changes from protected baseline files.
Final incumbent C4 request76 exhausted2,048tokens without visible output	Retained failure and native usage; no retry or replacement. Both incumbent C4 arms and overall ranking withheld. Final fresh restoration checks cover the runner’s skipped post-block assertion.
Offline GPU aggregation initially lacked per-window coverage and analyzer binding	Independent review required per-window sample counts, duplicate-series rejection and source/analyzer hashes; fixed with a synthetic missing-window/duplicate regression check. v1 remains retained as superseded.
Evidence negative test r2 had a misplaced test block	Original NameError log retained; corrected test layout. Five tests pass under optimized Python in evidence-checks-r4.log.