Eve Task 13.4 — Resource-Safe Load and Bounded-Soak Report

Observed 2026-09-14T16:39:42.097Z; implementation c6396240983aafd04e40168aa130e44014016031; immutable record snapshot ca2ff9b34727785d9d7483806e9fd8b0df3e58c1; focused review repair e3e774b15c6f4dac07d2157a3de699e2e544ff6a; supplemental receipt snapshot 3b5f71d7e1012539851444b7a1ade7d8f61afeb2. The preregistered contract digest is ae44527017c95e916918e8c1d383d91562f08f918eb9455acd5c04a04e2088a6 and the retained record digest is 044a455244fe2d08de0b7d5cc005f9aae4080dfbc6ff22d4c7ed395f80f6331c.

Result#

PASS. Five expensive surfaces ran serially for 300.265 seconds total with 900 observations. Maximum simultaneous expensive surfaces was 1; the host-safety abort flag remained false. Every observation succeeded, no observation timed out, and no duplicate effect occurred.

Queue and watcher deliberately replay every operation once, so their 100% duplicate-attempt rates prove idempotency pressure; the outcome guard is the 0% duplicate-effect rate.

Stream, tool, and provider result#

The production OpenRouter resolver served deepseek/deepseek-v4-flash-0731 through fp8 routing. All 100 turns executed the required read-only tool exactly once and completed. Total-turn latency was p50 1534.783 ms, p95 2711.646 ms, and p99 3201.943 ms. Time to first meaningful tool/text event was p50 764.959 ms, p95 1222.470 ms, and p99 1420.187 ms.

Tool-bearing first iterations were narrowed to baidu/fp8. The final permitted iteration deliberately disables tools; 94 turns remained on Baidu and six accumulated Baidu plus StreamLake response provenance across their two iterations. This is consistent with the request-scoped pin and does not claim StreamLake served a tool-bearing request.

The canonical prompt-template descriptor SHA-256 is 3038d890424ad07a093b383e436ff4545f00c32a2a31afdc0331b9661b6ec0e3. It represents the source-fixed system prompt, token placeholder, tool schema/result shape, iteration policy, and output ceiling. Per-observation aggregate input/output tokens and OpenRouter-reported billed cost are retained.

Actual billed provider cost was $0.023782300 total. Per outcome: p50 $0.000237224, p95 $0.000252624, p99 $0.000253944, maximum $0.000254824. All 100 outcomes reported cost across 63052 aggregate input and 5252 aggregate output tokens, an observed blended $0.348183/M reported tokens.

Cost reconciliation: the registry's 2026-09-05 endpoint-feed values ($0.04998/M input and $0.09996/M output) were an admission-time planning snapshot, not the billing formula for this 2026-09-14 run. Applying them to the retained aggregate tokens predicts $0.003676329, while OpenRouter's retained usage.cost totals $0.023782300 (6.469×). The receipt aggregates two iterations per turn but does not retain per-iteration provider price, cache-token class, or provider billing adjustments, so a defensible input/output-rate decomposition cannot be reconstructed. The evidence therefore treats the provider-reported billed-call total—not the older registry planning snapshot—as authoritative, and records the missing billing dimensions explicitly instead of presenting the two as equivalent.

Persistence and runtime provenance#

• Queue: PostgreSQL 16.14 (Debian 16.14-1.pgdg12+1), WorkbenchIntentStore through WorkbenchQueue, four tenant queues, restart readback passed.
• Vector search: Qdrant 1.19.1 (6ab21cac18ebb6f4ae29102c7f8f5cc11affd5de), 1,000 indexed 16-dimensional records through QdrantVectorBackend, restart readback passed. The focused real-Qdrant 1.19.1 supplement then proved two distinct namespaces, fresh v1→v2 update visibility with stale absence, zero cross-namespace records, a new-adapter readback, point deletion with peer preservation, and both collection deletions returning 404 and rejecting adapter queries.
• Watcher: PostgreSQL 16.14 (Debian 16.14-1.pgdg12+1), PostgresWatcherStore plus watcher-engine, 100 deliveries across four watchers, restart readback passed.
• Long job: ContentPipelineService with FileContentRunStore and FileCheckpointStore; 100/100 runs persisted and recovered after restart readback. The focused recovery supplement sustained 1118/1108 ms of work before/after a durable incomplete checkpoint, resumed it through a fresh PipelineRunner and FileCheckpointStore, preserved one run ID across two dispatches, and did not replay the already-completed step.

Resource envelope#

Process CPU is percentage of one logical core and may exceed 100% for parallel local work. The host has 4 logical CPUs.

The contract required at least 2.00 GiB available memory, at least 1.00 GiB on every critical disk target, and no more than 256.0 MiB process-RSS growth. Every surface remained inside all three guards.

Observed failure, diagnosis, and recovery#

Attempt 03 stopped honestly after the stream/tool surface: 93/100 observations failed, all as forced-tool-not-executed. The bounded endpoint diagnostic then measured:

• baidu/fp8: 4/4 exact tool executions, 0.0% errors.
• deepinfra/fp8: 0/4 exact tool executions, 100.0% errors.
• streamlake/fp8: 0/4 exact tool executions, 100.0% errors.

The production repair introduced a request-scoped toolOnly: ['baidu/fp8'] route while preserving the broader non-tool route and explicit environment override precedence. Attempt 04 then passed the same locked stream/tool workload 100/100 before the other four surfaces ran. The failed record, diagnostic, repair source, and passing record are all retained; thresholds were not weakened.

Verification cadence#

Recovered stage: CANDIDATE-REVIEW → FOCUSED-REPAIR after independent reviewers identified four semantic gaps. Reused green receipts were shared-AI lint, BFF lint, 41 shared-AI tests, 43 BFF registry/config tests, shared-AI typecheck, BFF typecheck, and all five Attempt 04 load surfaces. Review invalidated only Qdrant lifecycle completeness, incomplete-job resume realism, price reconciliation, and exact-one-tool verifier enforcement. Repairs added direct fail-closed controls plus one focused Qdrant lifecycle run and one focused durable-resume run. Attempt 04 and its provider, queue, vector-load, watcher, and original long-job surfaces were not rerun.

Limits#

• This is a supervised local candidate with synthetic minimized inputs, not production traffic or a rolling 30-day SLO window.
• Each surface receives sixty seconds of observation. This can reveal immediate saturation, leaks, duplicate effects, and deadline failures, but it is not a multi-hour or multi-day endurance claim.
• The stream/tool surface uses the real production OpenRouter binding and records billed cost; queue, vector, watcher, and long-job surfaces are local dependencies with zero variable provider spend.
• The Qdrant surface proves the shipped Qdrant adapter. The local pgvector cosine operator is not claimed because the host extension currently fails with recursive cosine_distance stack exhaustion.
• The long-job surface exercises the production durable content pipeline with real filesystem run/checkpoint stores and deterministic local model doubles; it does not claim an external generation provider or deployed object store.
• Destructive dependency outages, network partitions, process crashes during effects, restore RPO/RTO, deployed dashboards/alerts, and a game day remain Tasks 13.5 through 13.7.