Conversation
Use immutable released v5.5.0 and frozen complete candidate 1c565ca, separately identify the evidence harness, archive all three exact sources before building, reject candidate overrides and repeated attempts, and retain original binaries/diagnostics/failure inputs. Keep all statistical thresholds, sample budgets, fixture mutations, measurements and existing reporting helpers unchanged. No product, VERSION, historical release or risk-disposition changes.
首次独立 ABBA 已触发既定阻塞：暂定 HOLD（2026-09-13）Performance #630/run34756138336 原始 attempt1 已完成；artifact10317871192（35,551 bytes）下载后SHA256与GitHub一致： 独立离线逐对重算：19×4共76对完整，72组有内部数据的完整counter/breakdown/hit-miss向量相等；4组external内部计数仍不可用，不填零。external timings-disabled-no-op的四对变化为：
配对中位差 +1.5423ms / +12.8314139%，同时超过事先声明的 +1ms 且 +10% 门槛。因此 #188 保持 Draft/HOLD，不能仅凭 workflow success 合并。 本PR产品未修改并不是豁免依据；没有重跑、剔除样本、AA扣除、改变门槛或已证明噪声的归因。 其他不利样本也保留：target-scale-cold中位差+369.7075ms/+8.9261485%，最大+3976.505ms/+67.2026138%；target-scale-no-op中位差+4.1785ms/+11.7533315%，最大+31.942ms/+92.6795300%；target-scale-no-op-j1四对全慢，中位差+4.462ms/+10.0651136%。全部19场景有利/不利原样入账。 累计v5.5.0→冻结候选的独立长矩阵仍在运行，不能拿其结果替代此次PR门槛。Windows native/自举包的已有首次运行继续收集，不重新提交head或制造新测量。最终完整证据与接手决定另在本PR/#164记录。 |
Iviesever
left a comment
Iviesever
left a comment
There was a problem hiding this comment.
实质验收决定：HOLD，保留 Draft，不合并；5.6.0 不放行
本轮实现并验证了累计比较入口，但本 PR 首次独立 ABBA 跨过预先声明的阻塞门槛。累计比较还发现独立的 private129 暖启动退化。Actions 的 success 表示采集/测试完成，不是全部人工性能门槛通过。没有重跑、换 head、调阈值、删样本、AA 扣除或合并豁免。
精确对象与变更范围
PR base：1c565ca9f364b0bcb380565c4684ea543b449e04；head/harness：8e9c2882014c84feb8021bcba5151d383b1d39db；head tree：bf471635b82aeba40597b42d15871937dee267da。原 CI test-merge ff1ed49951869b3418e78ea4a4be474706aa40e1 的两个父提交及 tree 均精确核对。单提交、仅两个文件 +230/-50，产品、原生清单、VERSION 5.5.0、reporting helper、独立 ABBA harness 均不变。
累计实验的被测对象另外固定为：released v5.5.0 源码 08cdc20a9f9380e18132d21a619288224fd07fd4 / tree f6ecbed37e6c6caf135d18ec734955a266f491d3，对完整候选 1c565ca… / tree 96629b0d75061027ca0eacd4a5522806215b95f7。不是把本 PR harness 当作被测产品，也不是拿历史发布 exe 冒充本次重建程序。
迁移覆盖工作流与 Python 两层旧基线；构建前/测量前核验三套实际源码，构建前归档；拒绝错误候选、重复 attempt 与覆盖证据目录；各累计运行独立 concurrency group、取消关闭、唯一 run/attempt artifact；保留实际两端 exe、seed hash、完整构建 transcript、原始调用/诊断/失败输入。没有增加产品自动恢复。
本地 AST 比较确认原统计、计数、warm audit、夹具、源码变更函数完全相同，完整测量循环的原12条语句保持，仅增加完成性检查/结果状态；12组统计正反对照一致。身份自检含21个拒绝变体、原异常传播、失败输入/PDB排除、不覆盖检查。Linux 本地只做 portable/离线验证，不声称本地 MSVC 实跑。
一、首次 Windows 原生与包门槛
原始 attempt1：Native C++ #758/run34756138328、Native Release #635/run34756138329、Documentation #199/run34756138331 完成通过。
两配置四个配对 job 的完整原始日志逐项核实，各八子分片 8,9,9,9,10,12,11,11，各 79/79 个程序实际执行。Debug jobs：103721813045/103721813078/103721813032/103721813056；Release：103721584724/103721584710/103721584733/103721584715。prewrite44、recipe28、target84、batch50、domain39/map33、前台24检查与原始失败诊断保留；不将 Release 测试宿主一概当作实际 Release 目标证明。Debug prepare 本轮核对步骤状态；Release prepare job103720708391 已完整读原始日志，不混淆两者。
完整读取 self-host job103721880961：pinned seed → Stage0 → Stage1 → 清空状态 → Stage2，后两代各实际编译83个产品TU；七文件包、SHA256 sidecar、安装/重装/卸载/PATH 所有权通过。Stage1 3819c56b0149d7741926854aa2d7adfbe13987c2f2ce1f6d33fa762053a3d3ae，Stage2 703e86b34f556cc810d1c3eeb7f64eccc5900d98f8a2091f02af1d46a7d1077f；二者不同，不宣称逐字节可复现。
包 artifact10317163858 已下载独立校验：外ZIP 916d13b44d24e59c4bb447edf90b07632a5da913c0836100255a4645570afeeb；内ZIP/sidecar b712488c7986f73a573dd203e42bf358f87053051f0da815f3f8e8e05e928ae5；七文件扁平清单、VERSION5.5.0、AMD64程序精确等于已验证Stage1，其余文本与源文件仅有允许的CRLF差异。publish-release 明确 skipped，不计为发布成功。
二、本 PR 独立 Performance #630：明确未过门槛
run34756138336/job103720708367，artifact10317871192，SHA256 41312a0d5c55aa909c3941408da68b441a4ab3f4f4b493308304db53b36d97c4。原始 observer-OFF 19×4、76对完整，72组可用的完整 counter/breakdown/hit-miss 向量一致；external缺失数据仍为不可用。
external/no-op 四对原始变化：−4.1381ms/−25.4633504%；+4.7091ms/+39.7365579%；+2.4861ms/+21.0086448%；+0.5985ms/+4.6541829%。配对中位差 +1.5423ms/+12.8314139%，同时 >1ms 且 >10%，触发既定 PR 阻塞。 产品零改动不是豁免理由。
其余18场景同样原样保留：target-scale-cold 中位+369.7075ms/+8.9261485%、最不利+3976.505ms/+67.2026138%；target-scale-no-op 中位+4.1785ms/+11.7533315%；target-scale-no-op-j1 四对全慢，中位+4.462ms/+10.0651136%。不将这些同因归到噪声、缓存或OS。
重要证据边界：现有独立ABBA原产物只保存比较JSON及Actions日志，没有归档这次实际重建的两端exe/hash。累计实验或self-host的exe不是它们的替代品；以后重建也不能补造历史身份。此缺口限制因果诊断，不使本次失败门槛失效。
三、首次累计 v5.5.0 → 完整候选：采集完整，但有发布阻塞
Release Cumulative Evidence #2/run34756138330/job103720708314，artifact10317617795，SHA256 d70a0478c8b060e1da1b64e1361ec9ee69978277eed3d2c671149cff372c4530。
独立重算全部源码Git树：baseline389文件、candidate438文件、harness438文件，与上述精确tree一致；harness只比候选改两个入口文件。归档、preflight/measurement/final身份与run/attempt相符。实际build transcript验证相同pinned seed（SHA256 f58d78066a2ec6c4d174e39fcb9c748776cdc4a6137a0d5a592a216b3eba6dd0），分别真实编译v5.5.0的80TU与候选83TU，均Release/VERSION5.5.0。两端构建脚本、seed脚本和reporting helper源码相同。
实际被测二进制归档并验hash：baseline 46a3fc9fd4826b16f5130895438a82601d6531a29000982720f88cfc94fcc95f；candidate deab6c9dea5342f3a83ca79e2be3f51d483b03a8f045fbfa8411a8e6ad645c8c。
20场景/584评分对/1446条MQB调用记录完整；1168评分调用timings-OFF，另258次instrumented correctness audits。逐条核对原stdout/stderr及哈希、退出码、pair/call对应、实际AB/BA顺序、共享夹具、同recipe、实际Release参数、warm前后状态与重建范围；258份audit的129组完整计数/缓存向量也相等。1444次exit0、2次故意编译失败exit4，原诊断与随后调用者修复保留。独立离线重算20行全部bootstrap/配对统计与原结果一致。
唯一累计practical flag：private-default-jauto（private129暖启动），40对中26对更慢。 配对中位差 +3.4128ms/+6.8036063%；原bootstrap区间 [+0.11555,+5.10205]ms；原practical阈值 1.5527055ms。中位数超过阈值且区间下界>0，明确记作待处置发布阻塞。原40/20/6预算、5000次bootstrap、seed5500、阈值max(暖0.5ms/重建及冷5ms,base中位数3%)全部未改。不能用#630单PR门槛代替累计规则。
common129冷构建6对中位+69.6847ms/+1.4245237%，未触发该行practical flag，但原配对1 4950.8644→9831.1244ms（+4880.26ms）、配对4 **5862.4642→9404.6717ms（+3542.2075ms）**完整保留；另外两对有利样本−1167.5053ms/−269.1931ms亦保留。中位数/flag不能清除秒级尾部，更未证明与历史common129同因。
边界：private129前后独立instrumented audits都为129 compile hits、1 link hit、0工具/缓存写入，计数向量相等。但它们不是40对评分样本本身，不能将其内部work移植成timings-OFF退化原因。12次直接程序smoke检查由原脚本完成性断言支持，无逐次原始recorder/二进制归属记录，不扩大成12份各版本独立产物证明。本轮不存在可用于定位评分冷尾的OS/失败时刻工具跟踪。
四、下一接手点与历史账
本 PR 保持 Draft/HOLD；main仍以1c565ca为预期起点，不转Ready、不合并、不更换head、不重跑这两组矩阵。 两个问题分账：#630阻塞本PR验收；累计private129和冷尾阻塞完整候选发布，不能相互替代或豁免。
下一单主题：对上述原始外部elapsed退化做有界归因与证据补强，先利用已保存的累计实际exe、源树、调用与工作量信息，明确旧#630缺失的历史二进制/OS证据不可补造。任何新增诊断实验先声明假设、固定预算、原始对照、工具/环境/两端exe身份与观察者影响，作为新增诊断证据，不作为同head重跑替换旧样本。找到具体产品问题再另拆修复PR；不能把本轮直接改称“环境噪声”后合并，也不能重建一对新exe说成#630原exe。
#172继续Draft/HOLD；历史#172/#176 C1041、#170 +1.6946ms/+13.7988%、#180#599 +2.07175ms/+16.2700%、common129 +5.354s/+6.480s、#625 +63.0149ms/+392.6039687%及#627不利尾部全部保留。无发布风险清账、无默认CLI取消/写锁/完整事务/自动恢复/会话/驻留扩展、无历史tag/资产修改。
证据包保留三份原始Actions ZIP、原样测量JSON/诊断/源码/实际累计exe、独立审计及本地验证记录。离线ABBA审计器在旧#187控制数据上曾因百分比显示舍入精度假设错误失败，修正仅对齐两位显示精度；原代码/失败日志与修正记录均保留，未变原始样本公式或门槛。一次Debug job ID转录错误导致404，随后按真实job列表纠正，不算Windows失败也不隐去。实际#630、累计与补充审计各自首次通过的是证据一致性审计，不等于性能通过。
…aries Accept diagnostic-only changes after substantive review5190963120 of exact base1c565ca/head8c8d213/tree4fc13a. First Windows diagnostic318calls, Debug79/Release79, self-host Stage0/1/2, exact package/installer/docs, and original ABBA#631 passed their declared gates. Retain all original and new adverse evidence. No product change, no claim that old performance regressions or C1041 are fixed. #188/#172 and v5.6 release remain HOLD; VERSION5.5.0, default cancellation/write-lock and v6 route unchanged. Metadata-only performance#632/#633 skipped after validated first evidence; no duplicate measurement.
Accept only the four diagnostic files after substantive review5193446514 of exact base7526686/head4977036/tree1544b00. Original Windows Debug79/Release79, clean self-host/package/installer/docs, bounded78-call20-trace diagnosis and original independent ABBA#634 reviewed. External no-op +1.21235ms/+9.8760361423% does not cross the predeclared conjunctive gate; all slower pairs and link-only +164.53ms/+300.17% remain risk evidence. The baseline8.125s cl failure-control lifetime is a lead, not a root-cause or repair claim. #188/#172 and all historical C1041/performance release holds remain. VERSION5.5.0, product code, release assets and v6 route unchanged; no publication or repeated scoring.
Accept only three diagnostic files after substantive review5195682463 on exact basef359f519/head2aa3cc7/tree5f48fe9. First-attempt hosted replay, all Windows Debug79/Release79 programs, clean Stage0/1/2 package and installer/docs validated. Original independent ABBA#637 retains76pairs/72full vectors; external no-op paired median -0.11935ms/-0.965444662572195% does not trigger unchanged conjunctive gate, but all adverse discovery/modules tails remain. Recovered existing raw process exits/cycles/IO/fault counts without new diagnostic MQB calls or ETW captures; original18MQB captures,1050cl/16link and calibration0/17/0 reproduced. CPUCycleCount is not CPU milliseconds or wait-cause proof; historical failures and release risks remain unresolved. #188/#172 HOLD, VERSION5.5.0, product/old recorder/statistics/tags/assets unchanged, no5.6 publication or6.0 route change.
Iviesever
left a comment
Iviesever
left a comment
There was a problem hiding this comment.
2026-09-17 续接实质验收完成：原累计数据完整；#188 与 5.6 发布仍 HOLD
接续 #164 comment5709658468，保留本 PR 原 review5190674100 和 comment5653210188。本轮直接验收已有产物，没有新增提交、竞争 PR、分支推进、Windows/MSVC 实验、ETW 捕获或 ABBA，也没有 rerun/cancel。本 COMMENT 是当前维护者连接的证据审查，不冒充另一名独立人工 reviewer 的 APPROVE。
精确身份与原始产物
- Released 5.5.0 源码：
08cdc20a9f9380e18132d21a619288224fd07fd4，treef6ecbed37e6c6caf135d18ec734955a266f491d3，389 文件。 - 冻结累计候选：
1c565ca9f364b0bcb380565c4684ea543b449e04，tree96629b0d75061027ca0eacd4a5522806215b95f7，438 文件。 - #188 head/harness：
8e9c2882014c84feb8021bcba5151d383b1d39db，treebf471635b82aeba40597b42d15871937dee267da，438 文件。原 test-mergeff1ed49951869b3418e78ea4a4be474706aa40e1的两父提交和同树已重新读取；不是实际合并。 - 写回前 main 仍
55f57a84ad938da10d0e28b4578cd1aef6d7f903，VERSION 5.5.0。其 cpp tree 与上述候选、harness 均为fbe5667b2e2cdc5cba3a80f7c7486f3ce7935e74；这不把旧测量提交或二进制改名为当前 main。
累计 run34756138330，number2 / attempt1；完整 job103720708314 已读。原 artifact10317617795：6,215,366 bytes、2,931 文件，SHA256 d70a0478c8b060e1da1b64e1361ec9ee69978277eed3d2c671149cff372c4530。三份内嵌 source ZIP 按实际字节重建 Git tree 全部匹配；候选到 harness 仅原两个文件改变，acquire_seed/build_mqb/compare_reporting 在三树中逐字节相同。没有修改换行、源归档或 install.bat。
同一个固定 seed 实际 clean Release 构建 baseline80 / candidate83 个 TU，原构建转录保留163条 compile、警告和完整构建诊断。这是两份源码重建程序，不是历史发布包 EXE。 实际 AMD64 PE 与前后哈希记录对应：
- baseline
46a3fc9fd4826b16f5130895438a82601d6531a29000982720f88cfc94fcc95f - candidate
deab6c9dea5342f3a83ca79e2be3f51d483b03a8f045fbfa8411a8e6ad645c8c
seed 有原验证记录和哈希，但 seed EXE 本体不在本累计 ZIP，不能冒称全部构建工具均已随此 ZIP 留存。
584 对不是只看 summary：本轮逐项核验
12 个 warm 场景各40对、4个 rebuild 各20对、4个 cold 各6对，共 20 场景 / 584 对 / 1,168 个 timings-OFF 评分调用。加上258个独立 instrumented audit、16个 prime/seal及4个故障/恢复调用，总计 1,446 条 MQB 调用、2,892 份 stdout/stderr，全部逐字节哈希对应原调用。
完整调用顺序与原采集循环一致，配对编号、AB/BA顺序、同一 cwd、除EXE外相同 argv、默认/verbose/-j 参数、评分OFF字段、实际编译/链接/归档范围及60条声明的源码修改hash均核对。两次故意 compiler failure 均真实exit4，原错误文本保留，后续恢复单独记账；没有用恢复覆盖失败。另12次产物运行在原脚本检查和完成标志中有记录，但没有逐次程序stdout/程序EXE归档，不能假称本轮独立重放它们。
258条审计JSON与原始timings行相同；其129组可用A/B cache/counters/breakdown向量一致，全部保持相应cache hit和无工具启动/无cache write。评分调用自身仍没有内部计数，不能把前后审计计数插值成每次评分的测量。 12组warm前后inventory相同；这些inventory只含路径、大小和mtime_ns，不是cache/产物全文哈希或任意并发写入的证明。
独立脚本按原运算顺序重算全部20场景的配对中位、百分比、MAD、nearest-rank P95，以及原 5000次、seed5500、索引124/4874 的配对中位重采样区间，结果与原JSON逐值完全相同。没有删除样本、扣AA/observer、修改阈值或用分组替代正式统计。
累计阻塞：private-default-jauto 必须实际处置
这个场景是129个编译单元、各自私有头文件的正常公开 build/no-op 配方，不是此前显式 /Zi /Fd 的研究夹具，不能借那个范围结论排除它。
- 40对，26对候选更慢、14对更快。
- 配对差值中位 +3.4128ms；配对百分比中位 +6.803606297454866%。
- baseline绝对中位51.75685ms、candidate绝对中位54.18715ms；两个绝对中位之差不是配对中位，不能混算。
- 原实用门槛
max(0.5ms, 3% baseline median)= 1.5527055ms。 - 原重采样区间 [+0.11555,+5.10205]ms；两条件同时满足，
practical_regression_flag=true。 - 全样本范围 −13.8871到+24.9711ms，MAD6.4171ms、nearest-rank P95+16.2685ms。正负样本均保留。
仅作事后描述、不改原门槛：AB和BA各20对，配对中位分别+3.1306/+3.6129ms，两组各13对更慢；前后各20对中位+2.8283/+4.3275ms。这不能证明独立性、具体根因或排除环境影响。
该场景四个前后审计均为129 compile hits+1 link hit、131次cache open/276646 bytes、550 snapshot requests/410 unique paths、3个background threads，cl/link/lib/cache-write均0。可用记录不支持简单改称“多编译或多写cache导致”，但计数相同也不能证明时间等价。审计work值有重叠，不相加或从外部elapsed相减归因。本轮没有定位这3.4ms的因果来源。
其他19项没有触发原标记，不等于所有风险解决：common-cold六对保留 +4880.26ms、+3542.2075ms 等原尾值以及−1167.5053ms；其配对中位仅+69.6847ms、区间跨零，不能借中位或未标记删除尾值。全20行及584对数据随离线交付保留。
#188 自身的独立 #630：另一个明确阻塞
run34756138336 attempt1 / artifact10317871192，35,551 bytes，SHA256 41312a0d5c55aa909c3941408da68b441a4ab3f4f4b493308304db53b36d97c4。19×4=76对逐对复算、72组完整可用counter/breakdown/hit-miss向量一致；4组外部null/-1保留。原JSON仅存比较数据，不含这次实际A/B EXE；上述累计EXE和其他安装包都不能补称#630测量身份。
外部四对差 −4.1381、+4.7091、+2.4861、+0.5985ms；配对中位+1.5423ms / +12.831413867442535%，同时超过原>1ms且>10%门槛。collection success不等于性能通过，没有产品改动也不是豁免。另保留target-scale-cold中位+369.7075ms/尾+3976.505ms、scale-no-op与j1不利值及全部有利值。四处.NET/Python中点显示舍入差异单独记录；门槛始终使用原未舍入值。
验收边界、决定与下一接手
本轮重读 #188 五条工作流完成状态及 main Native#764七个job/step成功；原Native79/79、selfhost、包/安装的详细验收仍指向旧review5190674100。本轮完整读的是累计job和两份原产物，没有重新执行Windows测试，也没有重新读完所有原生/安装日志。
本地只读审计新增的两个拒绝均保留：一处额外argv检查错误地省略了原discovery的main.cpp/--env vs；一处将原PowerShell百分比的先除后乘改写为先乘后除造成末位差异。依照未改原脚本修正检查器后完整通过；首次成功的全部统计与最终增强校验输出仍逐字节一致，没有更改任何原评分或接受标准。
数据完整性/原统计复算：通过。合并#188、发布5.6.0：HOLD。 #192/#641、#172及历史不利证据继续独立保留。当前head保持Draft，不转Ready、不合并、不发布。
下一单主题从main55f57a84接手 private129正常no-op的退化归因与最小处置，不再把“重新验收#188”或“再跑RM双场景”作为下一轮。 已保存实际原A/B EXE、全部原评分/审计与baseline→candidate关键源码diff。先核对Application/普通target进入run_impl<false>、TargetCompileWave检查与共享freshness、link inspection的真实可达变化和代码生成；现有原scheduler run本体未因新增admission-stop函数被证明变慢，不能先把取消分支判为原因。只有可验证原因才做单主题最小修复/回退，并保持freshness/完整诊断；需要Windows判别测量时另行登记具体假设与独立固定预算，不重跑旧584/76取绿或重复使用已耗尽入口。#630与#641不能由private修复自动抵消。6.0长期路线、5.6功能冻结与最终独立VERSION/release顺序不变。
First run35191361133 failed before download or any MQB call. Preserve artifact10484325806 and explicit correction preregistration #164 comment5710273155. Windows ZipInfo normalized the self-test's backslash name; test literal invalid member names directly against the same validation, not an already normalized round trip. Only the new corrected head's run_number2/attempt1/exact parent may consume the remaining original58 calls. Fetch three commits for the unchanged main diff check. Original sample schedule, A/B files, fixture, warm checks, metrics formulas, no-retry and stop rules unchanged. A redundant ABI assertion remains before any MQB call; local source matches uploaded bytes and all tests passed, including reproduction of old failure with Windows separator semantics and both separators after correction. These local models are not Windows measurements. No product changes, extra MQB builds, ETW, PR, main update or acceptance-threshold changes. A further failure stops; no third execution authorization. #188/#630/#192/#641 and 5.6 release HOLD persist.
1 participant
Scope: cumulative evidence entry only, not a product optimization or release
Continue #164 comment5653085411. Cumulative correctness mapping is closed; this PR changes only the entry/workflow and provenance/failure retention of the existing cumulative comparison. The initial
perf:title invokes the existing independent per-PR ABBA workflow; it is not an optimization claim.Exact PR base
1c565ca9f364b0bcb380565c4684ea543b449e04; head8e9c2882014c84feb8021bcba5151d383b1d39db; treebf471635b82aeba40597b42d15871937dee267da. Exactly two modified files, +230/-50. Product, cpp manifest, native inventory, VERSION 5.5.0, reporting recorder/fixture helper, original independent benchmark scripts and historical tags/assets are unchanged. main #757 attempt1 previously completed; only #172 was open before this branch. Keep Draft until substantive review.Three deliberately separate identities
08cdc20a9f9380e18132d21a619288224fd07fd4, treef6ecbed37e6c6caf135d18ec734955a266f491d3.1c565ca9f364b0bcb380565c4684ea543b449e04, tree96629b0d75061027ca0eacd4a5522806215b95f7.Both unmodified release-source trees are built using one pinned historical seed and the same runner/toolchain context. These are freshly rebuilt binaries, NOT the historical release executable. Preflight saves all three Git source archives, their hashes, expected identities, run/attempt, sample policy and Python/runner identity before any build. Full build transcript, seed hash, both actual rebuilt executables, original calls/raw diagnostics and partial failures are uploaded. Final identities are separate from, and do not replace, the initial plan.
Each cumulative run has its own concurrency group,
cancel-in-progress: false, unique run/attempt artifact name, and rejection of attempt>1. Existing evidence outputs are refused. First unexpected failures remain failures; source inputs are preserved before temporary fixture cleanup without opening/copying PDB/IDB. Existing recorder retains failing completed calls or available timeout diagnostics. No claim of durable transaction, PDB quiescence or complete recovery.Gates declared before the first Windows measurements
Portable validation completed before publication
Reconstructed the original438-file archive to exact base tree
96629b0d75061027ca0eacd4a5522806215b95f7; remote proposed tree exactly matches localbf471635b82aeba40597b42d15871937dee267da. Python syntax/self-tests and diff whitespace passed. Independent audit found AST-identical comparison/counts/audit_warm/write_later/build_fixtures functions, identical results on12 favorable/adverse/threshold statistical datasets, correct YAML identity/concurrency/archive contract, and actual synthetic Git preflight/archive roundtrip plus dirty-source rejection. Statistical budget recomputes584 pairs /1446 calls. Environment is Linux; no local MSVC run is claimed.Unchanged release blockers
#172 stays Draft/HOLD. Historical #172/#176 C1041; #170 +1.6946ms/+13.7988%; #180 Performance#599 +2.07175ms/+16.2700%; common129 +5.354s/+6.480s; #186/#625 +63.0149ms/+392.6039687% and #627 adverse tails (including target-scale-cold+4751.707ms/+120.6877748%) remain. No noise/root-cause claim and no new green clears old evidence. No default CLI cancellation/write-lock, automatic recovery, full transactions, sessions or residency. Keep VERSION5.5.0 and #164's v6.0 route. Independent VERSION/release PR and final publication require later authorization.