docs(authority): record formal SQLite growth qualification failure

[loopx-agent]

Goal And Delivered Outcome

The existing authority RFC checkpoint lacks a current, full 64 KiB 10k/100k qualification result. Update its English and Chinese read-cost section with the unchanged formal runner's 13 passed / 1 failed / 11 missing ledger, source/runtime, key p95 values and remaining qualification boundaries. The warm-head growth result is 2.059x against the frozen 2x budget; its passing absolute latency does not cancel that failure.

Related to D2 #4224; intended base: main. This documents a qualification gap and its next diagnostic boundary; it does not close D2 or switch defaults.

Author Declaration

• Written by: model_agent — GPT-6, OpenAI.
• Specification: shared Goal authority RFC D2 and its existing read-cost checkpoint at 9ba6148fccf335f2e38d800812dbb398008229c6; overall roadmap R5.

Criterion	Disposition	Path / evidence
Preserve frozen budget outcomes and source/runtime applicability	implemented	Both authority ledger translations; formal runner at 5f51559dc
Separate storage-axis observation from full D2/default acceptance	implemented	All eleven missing rows remain disclosed; B remains open
Full D2 sustained workload, platform and recovery qualification	deferred	Existing D2 program and owner; obtain completed soak evidence and applicability before a replacement run

Scope And Continuation

Only two existing RFC ledger files change. The bounded next work diagnoses continuity-query, proof/codec, connection and host costs before selecting a repair, preserving hole/digest/receipt verification. The root cause is unproven. The existing D2 owner retains soak qualification. Complete within this documentation scope; runtime repair, release defaults, File comparison and supported upgrades remain independent work.

Future-facing pass: reuse the existing checkpoint and runner; add no duplicate runner, schema, test scaffold or Python decision owner. No live authority or optional capability change.

Validation

• Documentation revision: 64dcf238efb764554a649bf69fcb9dd8f98fcad8.
• Formal measurement source: 5f51559dc93b43d023a9c915c52ec1dc409e3e65, not the documentation head.
• Run state: finished. Inputs: synthetic isolated histories and public source scan.

Check kind	Result	Public-safe evidence / limitation
real_entrypoint / real_backend	failed	Unmodified examples/coordination/sqlite-capacity.ts --profile matched-64k --cli on macOS arm64 / Node 24.21.0 / SQLite 3.53.4 with explicit checkout Python and non-empty source scan. Both histories completed; warm head p95 2.360→4.860 ms, growth 2.059x > 2x. Source stable; both temporary stores cleaned.
manual	passed	Independently checked original report counts, both complete histories, unchanged scan digest and bilingual numerical/qualification parity. Raw local evidence stays private; the repository runner is the safe reproducer.
static	passed	git diff --check and native public-boundary scan over both changed files.
integration	passed	Native risk-based premerge selected 18 catalog/docs canaries and one boundary scan; no check failures. The first overall gate rejected a stale prior change-quality receipt; exact current-diff qualification was then recorded and the gate rechecked.
real_backend	not_run	Full eight-agent/four-writer workload, steady RSS, large-history recovery, consumer lag, upgrade/rollback, elapsed soak, runtime matrix, promotion, 1 MiB, 300k and 60-second burst remain missing in this profile.

Coverage: the documentation describes the actual failed result rather than claiming a passing runtime qualification. No implementation, thresholds, first-screen UI, optional capability activation or installed behavior changes. Raw traces, private state, local paths and credentials are excluded.

Frontend / Visual Evidence

UI impact: none; existing RFC prose/table content only. No frontend, Lark or CLI behavior changes; no runtime tests added for documentation text.

Type And Area

Documentation update — public authority RFC / R5-D2 qualification checkpoint.

Boundary Checklist

• Public-safe aggregate evidence only; no raw logs, private state, local paths or credentials.
• Existing D2 program reused without taking over another owner's soak.
• Scoped to the measured qualification gap; no weakened acceptance or default change.
• UI impact none.
• DCO sign-off included.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; xhigh

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

Exact-head self-review of #5973 at 64dcf238efb764554a649bf69fcb9dd8f98fcad8.

动机

发布维护者在决定是否让新 Goal 默认使用 SQLite 时，需要知道当前容量验证究竟通过了什么。 原记录只说容量证据缺失；本次补入完整 10k／100k 测量，明确 head 读取虽低于 50 ms，增长仍超过 2 倍预算，避免把局部通过误当成默认切换已就绪。

读者现在能在同一验收记录里看到 13 项通过、1 项失败、11 项缺失，以及下一步应先定位成本来源。 本 PR 不修改运行代码、冻结预算或 provider 默认值，也不宣布 D2 验收通过。 head 增长失败的根因尚未证明；持续负载、恢复、升级回退、平台矩阵等缺失项仍需完成。

改动思路

复用现有双语 RFC 检查点，补齐当前事实，保留绝对预算、增长预算和完整发布验收的区别。此次有用增量是把容量未知转为可诊断的失败证据；运行规则和既有停止条件保持不变。

具体改动

关键内容讲解

规格依据：docs/architecture/rfcs/shared-goal-authority-state-provider-v0.md，不可变 revision 9ba6148fccf335f2e38d800812dbb398008229c6。7.2 的测量披露在本次记录中落实；D2 的完整资格仍 deferred，包括持续负载和独立的十日证据。现有 issue #4224 已有 soak 启动记录，先核实负责人报告与当前候选适用性。

两份既有文档 +60/-4；生产代码、测试、runner、冻结阈值和自动加载指令均未改动。完整读取 PR body/diff 和原始报告，核对双语数字：源 5f51559dc93b43d023a9c915c52ec1dc409e3e65 的 unchanged matched-64k 正式运行是 13 passed / 1 failed / 11 missing。测量源与当前文档 head 明确区分。

反例检查：100k head p95 4.860 ms 低于 50 ms，仍因增长 2.059x 超过 2x 而失败；文本两处明确禁止用绝对通过抵消增长失败，也没有把 soak 启动认作完成。根因尚未证明；B 保持 open，不据此声称发行回归。

对主干的风险

主要风险是错误的发布结论，而非此 diff 的运行副作用。现有完整报告、失败行和缺失项相互一致；CLI/App 操作和长期调度、恢复语义未变。19/19 原生 premerge 检查通过，无失败或 manual hold，当前 diff 的 change-quality 回执有效；git diff --check 和公开／私有边界扫描干净。最初 stale receipt 拒绝保留，随后生成当前 diff 回执后通过，没有绕过门槛。按原生策略使用本地验证，未查询或等待 GitHub CI。

原始容量实验本身 FAIL，其中真实 CLI、head/receipt 校验及清理完成；它不覆盖完整多 agent 持续负载、稳定 RSS、大历史恢复、升级回退、平台矩阵、promotion、1 MiB、300k 或 burst。文档检查通过不改变这些限制。若数字被新证据纠正，可独立回退／替换这两处检查点；不得放宽冻结预算来改判旧结果。

我的整体评价

交付判断为 justified_increment：长期执行（long_horizon）保持，维护者读回体验（user_experience）改善；上述缺失资格仍未完成。未发现本 PR 的可行动阻塞项。批准这一份可独立审查、回退的事实记录增量；批准不代表 SQLite 默认切换或 D2 验收通过。相邻重构判断已做：此次复用现有 runner／检查点，不增加结构；后续由既有 SQLite owner 分离连续性查询、proof/codec、连接与 Host 成本后选修复。没有修改核心规则来追求较小数字。

English verdict: APPROVE — exact head 64dcf238efb764554a649bf69fcb9dd8f98fcad8 accurately records the failed formal capacity result and remaining qualification gaps. No runtime/default/threshold change; local native checks passed. Full D2 and default promotion remain unqualified, and the warm-head growth root cause is unproven.