Status: 9.1.2 tip L6 closed on 2026-10-03 for Code 6ad2d6dc — both performance.yml and hermetic-integrations.yml archived below. Prior tip L6 pairs remain on b91462d3 / ee8f68ad / 91d34757. Enterprise GA remains unclaimed: TB-QUAL1 has no verifier receipt on this tip, and CAR is not certified.

Tip 2bd6a234 keeps that withhold. Standing-context admission is 9fb1cb79. CI 37103577874 succeeded, and its hermetic integrations artifact has top-level passed: true on s3-compat.json, headless-cdp.json, and opentelemetry.json. Performance Qualification 37103577761 produced the nine release reports and each top-level passed is true. The F-table gate on this tip scored 44 of 44 kernel files at or above 95% line coverage (content_digest.rs 100.00%, loop_checkpoint.rs 96.74%, lowest durable_memory.rs 95.00%). An earlier instrumented run stopped on a contended repository lock before scoring; the rerun is the score above. just layer-c-live-e2e with A3S_TEST_MODEL=boyue/bailian/deepseek-v4.1-flash printed LAYER_C_PASS model=boyue/bailian/deepseek-v4.1-flash and exited 0. The same route served deepseek-v4.1-flash on the deep end-to-end receipt. A run that left the recipe default boyue/deepseek-v4-flash did not: the agent-protocol write suite observed reasoning events and no write tool. The DM-PROD1 host pack on 3b2f72aa is unchanged and does not cover this admission. Terminal-Bench was not started. There is no v* tag on this tip.

Tip 9f521169 keeps that withhold. CI 37110650044 succeeded. Its hermetic integrations artifact has top-level passed: true on s3-compat.json, headless-cdp.json, and opentelemetry.json. Performance Qualification 37112230393 was dispatched on this commit and produced the nine release reports; each top-level passed is true. The F-table gate scored 44 of 44 kernel files at or above 95% line coverage on the first instrumented run (content_digest.rs 100.00%, loop_checkpoint.rs 96.74%, lowest durable_memory.rs 95.00%). just layer-c-live-e2e with A3S_TEST_MODEL=boyue/bailian/deepseek-v4.1-flash printed LAYER_C_PASS model=boyue/bailian/deepseek-v4.1-flash and exited 0. Every suite passed, with no retry. The clean deep end-to-end receipt is parent 5f5528f4; each case served deepseek-v4.1-flash. The DM-PROD1 host pack on 3b2f72aa does not cover this tip. Terminal-Bench was not started. There is no v* tag.

Changelog fold 439cc740 keeps that withhold. Kernel sources are unchanged since 9f521169. CI 37117714232 succeeded on attempt 2. Attempt 1 failed the Windows library test baseline_sees_a_mode_change_on_an_already_dirty_file with PathDelta { paths: [], incomplete: true }; the rerun is the score. Its hermetic integrations artifact has top-level passed: true on s3-compat.json, headless-cdp.json, and opentelemetry.json. Performance Qualification 37117721745 was dispatched on this commit and produced the nine release reports; each top-level passed is true. The F-table and Layer C were not re-run; those scores stay on 9f521169. The DM-PROD1 host pack on 3b2f72aa does not cover this fold. Terminal-Bench was not started. There is no v* tag.

This record is the human-readable companion to the machine-readable release profiles. It documents what was measured, what was deliberately excluded, and why the budgets are suitable regression ceilings rather than universal speed claims.

GitHub reported digests for the uploaded ZIP archives. The artifacts are retained for 30 days; the workflow also runs weekly and whenever a measured critical path changes, producing a refreshed independently downloadable record.

Run 37117721745 on commit 439cc740 produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate). CI run 37117714232 on the same commit archived S3-compat, controlled CDP, and OpenTelemetry reports with passed: true. These rows do not replace the 9.1.2 tip record below, and they do not claim Enterprise GA.

Run 37112230393 on commit 9f521169 produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate). CI run 37110650044 on the same commit archived S3-compat, controlled CDP, and OpenTelemetry reports with passed: true. These rows do not replace the 9.1.2 tip record below, and they do not claim Enterprise GA.

Run 37103577761 on commit 2bd6a234 produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate). CI run 37103577874 on the same commit archived S3-compat, controlled CDP, and OpenTelemetry reports with passed: true. These rows do not replace the 9.1.2 tip record below, and they do not claim Enterprise GA.

Run 37082677055 on commit 6ad2d6dc produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate). Run 37082696544 on the same commit produced S3-compat, controlled CDP, and OpenTelemetry reports with passed: true.

Run 36428840848 on commit 139c4680 (the 9.1.0 GA candidate line) produced all nine performance reports with passed: true; run 36428848277 produced the hermetic integrations reports with passed: true. These are the current-line L6 digests for the Enterprise GA board.

Run 36160896419 on commit b91462d3 produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate).

Run 36165517450 on commit b91462d3 produced S3-compat, controlled CDP, and OpenTelemetry reports with passed: true. Hermetic integrations is the CI job that calls hermetic-integrations.yml.

Run 35947891976 on commit fa0a92ca produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate).

Run 35947892108 on commit fa0a92ca produced S3-compat, controlled CDP, and OpenTelemetry reports with passed: true.

Run 35667548466 on commit c9e26504 produced all nine performance reports with passed: true (agent-convergence, workspace-retrieval, workspace-retrieval-portable, flow-state-graph, code-intelligence, context-memory, durable-memory-semantic-refresh, persistence, evaluation-substrate).

Run 35667552718 on commit c9e26504 produced S3-compat, controlled CDP, and OpenTelemetry reports with passed: true.

Run 35639843682 on commit 9b28f066 produced all nine performance reports with passed: true after standalone CI stripped monorepo path patches and re-locked a3s-apofasi, a3s-sandbox, and a3s-vec.

Run 35642286881 on commit 515a50b0 produced S3-compat, controlled CDP, and OpenTelemetry reports with passed: true after aligning the hermetic access key with the in-repo s3-compat fixture (a3s-code-akid).

• Every timed profile used a Rust release build on an x86-64 Linux runner with four logical CPUs. Profiles that resolve /proc/cpuinfo reported an AMD EPYC 7763 64-Core Processor.
• Each percentile profile performed three warmups and 20 measured samples, except Workspace Retrieval, which performed 20 warmups and 100 measured samples.
• Provider and public-network latency was excluded. Workspace hybrid timing included authoritative source rereads from the warm operating-system cache. Durable semantic recall used a deterministic in-process embedding adapter, the reopened file-repository replay view, and SQLite reads from the warm operating-system cache.
• Context and memory corpus construction was measured and reported separately, but excluded from query latency. Code Intelligence workspace creation was excluded; its manifest scan, language-server process start, source read, and shutdown were measured explicitly.
• File persistence included JSON serialization, filesystem I/O, and fsync. The durable-memory profile separately measured source seeding, source reopen, each refresh phase, and semantic-query percentiles. It closed and reopened every repository/index/session owner but did not restart the operating-system process. One-off refresh timings are observations, not percentile claims. Flow/State Graph timing was an in-process CPU profile and excluded filesystem I/O.
• The convergence client was scripted. Its call, tool, token, and recovery counts are deterministic work-amplification gates; its sub-millisecond-scale fixture timings are not a model-latency claim.

All values below are milliseconds. A value written as p50 / p95 / max comes directly from the retained JSON report.

The file-save result is intentionally reported rather than normalized away. Hosted-runner fsync timing varies substantially across otherwise successful runs, so the contract uses a user-visible ceiling with headroom instead of publishing the fastest sample as an SLA.

The change-scoped evaluation-substrate-v1 profile now measures bounded evaluation-result publication, reopen/retention validation, and strict wire round trips without a provider or network. It is wired into the release workflow and increases the required report set from eight to nine. Hosted run 33844533910 produced all nine reports and passed the new profile on Linux x86-64 (four logical CPUs): result-write p95 137.600 ms, reopen/validation p95 0.865 ms, wire-round-trip p95 2.185 ms, 24 retained records, 15,511 persisted bytes, and two regular files. The profile intentionally excludes provider, network, and Cloud latency; those remain host qualification boundaries.

The change-scoped 2026-09-04 schema-v5 workspace-retrieval run qualifies the official a3s-vec FTS path alongside the existing A3S Memory semantic projection. The locked exact, identifier, CJK, hybrid, concurrent-replacement, and close gates pass. Native collection handles are flushed and closed before publication; a four-entry hot-handle cap plus transient cold opens keeps collection reopen cost bounded. Portable minimal builds pass the same lexical result contract and are identified in metadata as portable_bm25_v1. The schema-v5 benchmark separates its 25,000-record exact-vector workload from the four-file/512-chunk native hybrid workload. See the backend contract.

These tests establish Code-owned transport, parsing, lifecycle, and cleanup. They do not claim that a public search engine, arbitrary S3 deployment, or remote Collector will always be available or have the same latency.

The objectives correspond to user-visible boundaries and include substantial hosted-runner headroom:

• interactive local retrieval should stay below 100 ms at the locked corpus;
• local durable semantic recall should stay below 1 second at 10,000 active nodes and 384 dimensions, while stable refresh ticks must perform zero snapshot, provider, and publication work;
• assembling a bounded model context should stay below 500 ms;
• a 1,000-step graph projection or replay should stay below 2 seconds;
• a cold controlled language-server request and clean shutdown each have a 5-second ceiling;
• an atomic, synchronized approximately 1.25 MiB file snapshot should save within 1 second;
• memory ceilings prevent passing latency by retaining unbounded state.

Changing a workload, inclusion rule, or budget requires changing the emitted profile and this record together. A job timeout remains only a hung-process guard and must not be presented as product latency.

Dispatch the remote workflow so native compilation and release execution stay off the developer workstation:

gh workflow run performance.yml --repo A3S-Lab/Code
gh run watch <run-id> --repo A3S-Lab/Code --exit-status
gh run download <run-id> --repo A3S-Lab/Code

The current workflow requires exactly nine JSON files (including the native a3s-vec workspace and evaluation-substrate reports) and rejects any report whose top-level passed value is not true. The authoritative run above contains all nine reports, including the native workspace and evaluation substrate records; its archive digest can be checked before the 30-day retention window expires. The capability ledger explains how these profiles combine with deterministic correctness, SDK runtime, and external qualification evidence.