[DavidWallstructurallaw]

P4-W08 checkpoint

W08 activates the public in-memory audit_bundle interface. Published head 7102222dc62461ce44d3a32c46f1a947b4baea59, tree 3cc336fc224069d8211becbe902e3441feeb8f44. PR remains draft; version remains 0.1.0.dev0.

• The W08 implementation changes exactly 18 authorized paths: P02/P11/P12, V03, T01/T02/T03/T05/T06/T07/T08, N06/N11/N12/N14 and D04/D05/D08. A final W08 successor updates D04/D05 evidence only. All other 232 tested source hashes, including README/package metadata, remain unchanged after the passing run.
• Seven approved exports and the exact public signatures are available. None, {} and {"raw_file_digest": False} options are equivalent. Exact type checks reject unsupported or hostile options before dossier access, equality/hash/repr hooks or runtime creation. The API calls the existing same-invocation implementation without a second budget.
• AuditUsageError keeps its fixed safe code/message and normal traceback/context-manager propagation. Sixteen immutable report roots and two precomputed bytes methods remain unchanged. The entire audit_file function, CLI, analytical implementation, limits, goldens, schemas, guard, workflow and dependency pins retain their prior bytes.
• Old public-refusal tests now exercise actual public completion, validation diagnostics and usage errors while retaining private behavior, file/CLI refusals, effect observers and their negative controls. Original test identities and exact migration dispositions are recorded. Field-by-field public closure remains W09 work.
• A clean committed-source focused run passed 231 independent tests and 145 separately counted subtest events, zero failures/errors/skips, exit 0, 335.298 JUnit seconds. All 234 source hashes were identical before and after. It covered the twelve W08 test files/helpers and three affected N01 controls. This was actual local Python 3.12.14 verification, not the final four-profile gate.
• Actual clean offline installs from direct wheel, sdist and rebuilt wheel each executed the public API with only the product distribution installed. All 48 installed module hashes match. Each returned 290,127 JSON bytes and 384,800 Markdown bytes, exactly matching the independent encoding model and each other; literal semantic, unavailable/zero-denominator, immutability and no-effect checks passed. Six real archive inventories and their members were reconciled. The existing preparation and private-analysis witnesses, including original hero stopping behavior, remain exercised.
• The old local preparation unittest deletes its temporary receipt after asserting it; its passing JUnit is preserved, while hosted mode retains that receipt. The three public installation paths have a retained actual runtime receipt. No missing local receipt, raw hosted archive or platform certification is invented.
• Two separately owned repairs precede W08. W07 repair c8b6df2fda2aa7c5fdd276b3c1ad277f9711a206 changes N12/D05 only to preserve the immutable-assignment test across Python exception-type differences. W02 repairs aa0e0b16af29fa5bda3bd63cb0c697e1b7cb8942 and 5de0c24a217ec15b7ae3457752580d3d568a4041 change N01/D05 only: the retained assertion uses verified activation history, keeping the old stage-7 refusals and checking actual stage-8 public behavior. Each repair's stage-7 targeted test passed; the final W08 run also passed the stage-8 branch.
• Earlier quick-check assertion and test-selection failures are retained with their actual corrections. A read-only review caught an exact-integer receipt-count issue; it was fixed with boolean negative controls before the passing run. Independent review of final code, tests, installed artifacts and source bindings found no remaining blocking findings.
• New Current verification 40273272046 is running in all four profiles. Final hosted results and raw-artifact acceptance are pending. The prior Windows 3.13 suite used 2,300.669 of 2,400 seconds; the new matrix's timing remains unverified and no time limit was increased.

W08 implementation and scoped local/installed verification are complete. W09 and W10 have not started. All 228 per-field public obligation rows/counts and prior evidence remain preserved. File/native/CLI audit execution, owner acceptance, merge and release remain at their approved later stages.

Preserved W07 checkpoint

W07 composes internal audit delivery under one original invocation budget, deadline and emergency reserve. Published head e9b6bfd3f65de31841606a7b1d44e0fe4cb3282d, tree 813157051432dadfba98388578da4a7b3859bc2c. PR remains draft.

• The W07 commit changes exactly P01-P10, N01/N02/N10/N11/N12 and D04/D05. _audit_value now composes preparation, original analysis, disclosure selection, frozen report assembly, JSON, Markdown, content consistency and immutable delivery on the same ledger. Existing analytical schedules, fixed limits, independent goldens, schemas, workflow, tool pins, version and module inventory remain unchanged.
• Actual new witness/member occurrences are reserved after final Finding deduplication. Original private object identity is retained locally, existing components are credited once, and extra occurrences consume the existing caps. Partial output never escapes. Rejected, stopped and cancelled paths retain their actual acceptance and cause; one same-budget diagnostic attempt remains terminal if delivery fails.
• Bounded native scalar passes and a same-invocation frozen-key ordering cache eliminate repeated work while charging actual passes. Content consistency validates decoded JSON scalar content; original independent byte tests establish canonical spelling. The cache does not cross invocations or alter analytical semantics.
• A genuine accepted acquisition case completes both formats and parity in 9,885,484 work units, with 293,616 JSON bytes and 388,746 Markdown bytes. The larger frozen reference can still hit the unchanged 10M cap and is tested for a real stopped outcome. No general guarantee that every finite case completes is claimed.
• Delivered reports expose 16 immutable roots and two immutable byte outputs. Generic serialization cannot export private delivery caches. AuditAbort/AuditCancelled retain fixed safe messages and permit normal traceback propagation without changing other fields. AuditUsageError and public entry activation remain W08 work.
• Unified ten-file regression: 6,275 passed, 1 failed, zero errors/skips, 489.114 JUnit seconds; all 232 source hashes stayed unchanged. The sole failure was an existing fixed-message assertion for AuditReport(). Four constructor-default lines restored that behavior; the original assertion was retained. Final targeted closure: 6 passed, zero failures/errors/skips, 15.528 JUnit seconds, all 232 hashes stable. The other 231 files match the unified run. This records scoped closure, not a new all-green 6,276-case rerun.
• Prior private-runtime regression: 146 passed, 172.840 JUnit seconds, with its earlier source scope explicitly recorded. Original private schedule preservation and the two later public-only cache lines were independently checked. Unicode equivalence covers 1,112,064 legal scalars, with all 2,048 surrogates rejected, against the final encoder source; this component proof is separate from whole-invocation budget evidence.
• A separately owned W03 repair is inherited before W07. Remote repair commits a996abcdbf606953c7a8f78565b632cd6bfc72f9 and 59bf1de37eb4fe334453777092d205d99ff23b84 change only P03/N02/D05 and retain activation stage 6. Strictly empty complete traces retain their original payload using the existing scoped-no-witness carrier, preserving unavailable/null/zero-denominator semantics. Final clean W03 checkpoint: 69 passed, 42.012 JUnit seconds, 230 stable source hashes. Its initial incorrect new assertion and 68 passed/1 failed run remain recorded.
• Earlier integration/resource/security failures and collection errors are retained with their actual scopes and fixes. No failing golden, source meaning, old budget limit or guard was relaxed. D04 retains all 228 public obligation rows and previous W02-W06 verification records; D05 preserves the complete inherited repair history.
• Independent read-only review found no remaining blockers. Exact uploaded blob SHAs, all three local/remote tree mappings, native commit objects, stage-7 history, approved 17-path scope and clean 232-file published checkout were verified.
• W07 Current verification 40267314989 subsequently passed in both Python 3.11 profiles and failed in both Python 3.13 profiles. The two original failure artifacts and logs were reviewed during W08: each had 10,068 passing cases, one failed immutable-assignment exception-type assertion, and 533 successful subtests. The actual assignment was refused. The separately owned W07 repair described above preserves that protection and accepts the supported interpreter's FrozenInstanceError. These original failures remain recorded; the new W08 matrix is pending.

At the W07 checkpoint, W08 had not started and public APIs still refused execution. W08 above now activates audit_bundle. All 228 per-field public obligation closures, file/CLI adapters, merge and release remain at their approved later stages.

Preserved W06 checkpoint

W06 implements exact bounded Markdown and independent JSON/Markdown parity from one selected frozen report. Published head e2cd1217ac7873c338d1d0503c0cdef259ce2f3c, tree 3a7207fc515a2bc529aea613b9a0b6dd5417a3e1. PR remains draft.

• The W06 commit changes only P05/P07/N05/N09/N10/D04/D05. Pretty-J, fixed 16-root layout, complete Result/Reason details, exact half-up fraction display, dynamic minimal fences and one final LF are implemented. Source strings remain encoded data; headings use checked tool-controlled tokens, Result IDs and registered field keys.
• Existing JSON behavior and all 19 prior escaping function/method ASTs are unchanged. Independent fixtures/goldens, report/schema, analysis semantics, public APIs, runtime/resources, limits, workflow/pins, version and module inventory are unchanged by W06.
• Real selected reference output matches independent JSON 421,307 bytes and Markdown 564,555 bytes exactly. A separate probe used the same real finalization ledger for both encoders: 7,022,265 renderer work units, 7,023,292 total including reserve/port setup. Its frozen document was prepared separately; this is not W07 whole-invocation feasibility or public delivery.
• Independent parsing verifies all root fields, complete Results, canonical Reason details and exact fractions. Omission controls reject dropped unknown time, reasons, counterevidence, denominator or qualifications. Security controls preserve literal backticks, HTML/images/links, controls/bidi and multilingual text inside safe blocks.
• Final five-file focused regression: 341 passed, zero failures/errors/skips, 34.750 seconds in JUnit, exit 0. All 230 source-file hashes were unchanged across execution: parity 30, rendering security 26, report resources 48, prior JSON 148, independent oracle 89.
• The 48 output-boundary tests retain all 38 W05 identities and add 10 Markdown controls. Actual fixed-counter remaining-space tests cover below/equal/above, fences/indent/final LF, interrupted output and no paired delivery after one format fails. They do not claim allocating a full 64MiB report.
• Separately inherited W01 repair 6cbe03b8f89ad7422a8fdb7ac25aa2cb48c4a0b8 changes V02/D05 only, activation stage 5. Two outdated inert-slot fixture path literals now target permanently protected io/output_directory.py; all original 213 ordered cases and assertions remain. Full V02 on the actual clean W01 checkpoint passed 213 tests, zero failures/errors/skips, 187.770 seconds; 228 source hashes were unchanged.
• Prior W05 hosted failure and all raw evidence remain recorded below. The repair's first local attempt had 184 passes/29 setup errors because W01 edits were uncommitted on a W05-owned checkout; the guard correctly rejected that scope. After the real W01 commit, the full clean-checkpoint run passed without overrides.
• N09's first run had 16 passes/10 new test assertion failures: whole-string counting missed a prefixed qualification canary. Exact output/parsing already matched. The assertion was corrected; no product or golden rewrite. All corrected security tests passed in the final unified run.
• Independent read-only review found no blockers. The published local/remote trees, separate owners, stage 6 history and clean 230-file checkout match. The verified repair was inherited after the W06 run without altering tested product/test bytes; only D04/D05 then recorded known facts.
• W06 Current verification 40212846768 subsequently completed successfully in all four profiles, confirmed through run/job API records during W07. This status update does not claim a new raw-artifact audit.

At the W06 checkpoint, W07 had not started. The W07 checkpoint above now records internal whole-invocation integration. All 228 public pipeline/rendering obligations remain pending; public API activation, merge and release await their approved stages.

Preserved W05 checkpoint

W05 implements exact JSON bytes for the frozen logical report and prospectively bounded encoding on the existing finalization ledger. Published head 98dd68b71f84dad3286355dfdb05d528f3bcfab1, tree 60327a132ca62f444a132f42542fdbcb411f5357. PR remains draft.

• Changed exactly nine approved paths: P02/P05/P06/P08/P09/N04/N10/D04/D05. The plan, P01 report/schema, independent goldens, all old tests, analytical semantics, public APIs, Markdown, runtime boundary, budgets, workflow/pins, version and module inventory are unchanged.
• J uses Unicode scalar key order, prescribed lowercase control escapes, UTF-8 for other scalars, exact portable integer/private number spelling, native array order and original rational components. JSON adds exactly one LF. There is no generic source/dataclass dump or custom default/repr dispatch.
• The real private reference outcome followed by W04 selection/assembly and W05 encoding matches the independent 421,307-byte JSON golden exactly. Its finite component port used 9,509,070 work units. This excludes the earlier private analysis and future Markdown/parity; W07 must still establish complete same-invocation delivery.
• The new restricted port attaches once to the existing authenticated finalizer, retaining its ledger, start/deadline and reserve. JSON/Markdown byte counters are separate, fixed at 67,108,864. Charges occur before emission; equality is allowed; the next byte stops with L14. Earlier causes and real input acceptance survive later failures, and cancellation remains separate.
• Final seven-file focused regression: 432 passed, zero failures/errors/skips, 9.770 seconds in JUnit, exit 0. All 228 source-file hashes were unchanged during execution. The run includes 148 JSON cases, 38 report-resource cases, 89 independent-oracle cases and 157 existing input/analysis resource/diagnostic cases.
• Real fixed-counter and remaining-space +/-1 tests cover the final LF, separate counters, mid-encoding work/time/failure/cancellation, no prefix return and port identity. They do not claim encoding a full 64MiB document.
• Independent read-only review found no blockers. All 45 prior resource and six prior diagnostic function ASTs remain unchanged; the existing output Protocol method moved unchanged to the narrower inherited encoding interface.
• The first new JSON test run had 147 passes and one erroneous redundant substring assertion. Correct exact bytes had already matched. The assertion was corrected to isolate a slash scalar; raw failure evidence is retained, with no product or golden change.
• Local and remote trees match exactly. The actual published owner/stage-5 history and clean 228-file checkout passed. Only D04/D05 evidence changed after the final focused execution.
• W05 Current verification 40192106767, attempt 1, finished failure in all four profiles. Each completed with 2 failed, 9,872 passed and 533 successful subtests, zero errors/skips/timeouts. Both failures were V02 inert-slot probes incorrectly targeting the newly active JSON module. All four original ZIPs were retained and their digests checked, each with 48 raw files; ordered 9,874-test collection and 228 source hashes agree across profiles. The separate W01 repair recorded above retains every assertion and test identity. New hosted success remains pending.

W06 was pending at the W05 checkpoint and is now implemented above. All 228 public pipeline/rendering obligations remain pending. Public APIs still refuse execution; whole-invocation report delivery, merge and release await their approved stages.

Preserved W04 checkpoint

W04 implements explicit, bounded disclosure selection from real private analytical outcomes and connects those selected fields to internal logical report assembly. Published head: dfbbbac7a187cbb7a37d775ea7fbc786b82d1110, tree f3c5754321a17c4a7ee6971fa13e406846394a78. PR remains draft.

• W04 changes only P03, P04, N02, N08, N16, D04 and D05. Native IDs, attribution, support, qualifiers, lifecycle, array order and duplicates, protected/shared/unknown relationships, and exact correction source slots are retained. Unselected excerpts, contact/display/identity fields, extensions and raw locators stay out. Required retained text may still identify a subject; this is not an anonymity or publication-safety guarantee.
• The selector uses explicit field tables and the caller's bounded port. Assembly accepts a retained correction component's already validated Finding home only after exact owner/scope matching. No schema, independent golden, analytical semantics, public API, rendering, budget, workflow, pins or module inventory changed.
• Final focused regression: 186 passed, zero failures/errors/skips, 73.047 seconds, across assembly, identity, disclosure security and independent-oracle files. All 226 source-file hashes stayed unchanged during execution. The later repair inheritance changed only V01/V02/D05; the tested W04 product/test/fixture bytes are identical in the published tree. Evidence-only D04/D05 updates followed execution.
• The separate W01-owned repair commit ea96496b43365e5c5f5f7f4869945be4f5ea085f changes only V01/V02/D05, retaining activation stage 3. It reuses a fresh read-only AST and import targets within one verification call, without cross-call caches or rule/budget changes. All 211 previous ordered V02 cases remain, with two new regressions. Its seven-file validation passed 384 tests plus 192 successful subtests, zero failures/errors/skips; all 224 source hashes were unchanged.
• Local three-run same-source guard measurements improved median time from 1.1700 to 0.5638 seconds. This does not prove that Windows CI now finishes within its limit.
• The two published trees exactly match their locally verified commits. Published history, separate ownership, activation stage 4, path scopes and the clean 226-file checkout were verified.
• W04 Current verification 40175369950 finished completed/success. All four Ubuntu/Windows and Python 3.11/3.13 jobs succeeded. This closes the earlier in-progress status using the actual GitHub run/jobs API; it does not claim a new complete raw-artifact audit. The legacy selected-toolchain workflow was skipped under its existing condition.
• This is an internal W04 checkpoint. All 228 public pipeline/rendering obligations remain pending. W05 was pending at that checkpoint and is now implemented above; public APIs still refuse execution; merge and release remain pending.

Earlier unsuccessful local runs, fixture corrections, the two actual product defects and the old Windows timeout remain in the evidence package. The phase plan SHA-256 remains 5fac7a2a796338faae9b98af6e0cc4519396e378ba10b8c5afeb4d029d97cd48.

Preserved W03 checkpoint

W03 implements bounded internal logical report assembly, stable semantic IDs and closed references on the approved Phase 4 branch. The exact published head is 95a8462105e5e5c5cfb3ab7f24cbf8b8cb37611f (tree 55e8522ef226481ed380f59b0623d39e4a3fc000).

• Only P01, P03, N02, N03, D04 and D05 changed from accepted W02 7919288.
• Local focused validation: 5,982 passed, zero failures/errors/skips, 70.78 seconds; 224 source-file hashes unchanged across execution. All 48 assembly and 16 identity cases passed.
• Full independent reference equality, real two-hop witnesses, compact two-Inquiry aggregation (30 capabilities / five domains), complete-body identity, cycles, ordered native arrays and reference/semantic negative controls are covered.
• Earlier failures remain recorded. The larger doubled reference really reached the unchanged 10M limit; its failure is not represented as a completed report.
• Actual committed-history, clean-checkout, path-scope and stage-3 package checks passed. The remote tree exactly matches the locally verified commit.
• The W03 run 40170942968, attempt 1, finished with failure. Three profiles succeeded; Windows 3.11 reached the unchanged 2,400-second suite timeout, with null exit code / incomplete / timed_out and no JUnit or final pass count. No partial count or active failing test is inferred. Windows 3.13 completed 9,653 tests plus 533 successful subtests in 2,397.33 seconds. The failed profile's 47 raw files and the Windows 3.13 comparison's 54 raw files were retained and inspected for the W04 diagnosis. The separately owned W01 repair below responds to this evidence; its new hosted result remains pending.

At the W03 checkpoint, public APIs still refused execution and W04+ remained pending. W04 is now implemented as recorded above. Renderers, same-invocation runtime integration and public delivery remain pending; merge and release have not occurred.

Preserved W02 history

P4-W02 验收完成。最终 head 7919288509e9a9232f5b26b004be5ba1d0a810f6，tree 82265150e952099466d7cbd5f49f03260a857538 的四环境完整 CI 全部成功，并已逐组核对原始工件。每组实际 9,589 tests，另 533 successful subtests，JUnit 合计 10,122，零失败、错误、跳过或不完整。最终提交归属 W01 的仓库准备修复，activation stage 仍为 2，交付任务为 W02。PR 保持 draft；W03-W10、merge 与 release 尚未启动。

公共报告此前没有正式闭合契约和独立字节预期。W02 增加不可变 AuditReport 声明、安全 transport、仓库 report schema，以及一个可完整运行私有核心案例的独立 JSON/Markdown 参考。现有分析语义保持，两个 public audit 入口仍立即拒绝；组装、披露、编码及公共执行留在后续工作单元。

已实现：

• 195 个 schema 定义、160 个闭合对象、586 个本地引用；16 根成员、57 字段绑定及 7 类值的合同。
• 全部 160 个对象位置和 627 个成员位置均有明确正例及缺失/额外/类型控制；原生状态、披露分支、数组元素、命名空间和局部关联有直接反例。
• 独立参考保留 78 Results、671 Checks、22 Findings、38 Reasons。原预期错误和逐源修正记录保留，没有用产品序列化输出来重录 golden。
• 所有 228 项公共流水线/渲染义务仍按实际未完成范围记为 pending，不能用类型或测试数代替公共功能完成。

验证与修复：

• 最初候选 05d51fce 的 run 38672908042, attempt 1 在四环境均失败：每组 27 failed、4,789 passed、另 533 successful subtests。失败身份完全一致，来自既有 V02 当前源码探针仍假定 stage 1：1 项冻结字节、8 项模块边界、18 项效果/依赖。四组均无超时、错误或跳过。

• W01 归属修复 d0121946c9a13a9b035d74921a7ac99c3ce0389f 只改 V02 与共享 D05。它读取已验证历史的实际阶段，并保留原 48 模块锚点、当前冻结字节及显式阶段反例。干净同树本地检查点通过全部 210 个 V02 测试，包括三种无 Git 收集模式；owner W01、activation stage 2。

• W02 后续提交补齐独立复核发现的嵌套覆盖缺口。最终 8 文件组合通过 6,217 tests 和另 12 successful subtests，83.33 秒，零失败/错误/跳过；N01 的 5,725 个测试全部实际执行。只读 schema 解码复用前后，6,217 个有序测试身份完全相同，测试输入和断言不变，schema 值/原字节前后校验通过。

• 第二轮本地检查点 23a07fd6a1fc135a82ee79ca4128586c09cb2af2 与当时远端候选 3b9d8a9 的 tree 相同。实际发布提交的历史、clean checkout、48 模块与 workflow 边界检查通过。

• 修订候选的 Current verification run 38677156662, attempt 1 已结束且整体 failure。Ubuntu 3.11 / 3.13 和 Windows 3.13 各 9,588 tests + 533 successful subtests，并已独立对账；Windows 3.11 的 pytest 在 2,400 秒处 TimeoutExpired，退出记录为 null/incomplete/timed_out，没有 JUnit 或末尾通过汇总，不能推导 partial 通过数量。全部四个原始工件保留。成功组的逐项耗时表明 V02 临时 Git 仓库用例有显著准备开销；这不证明 Windows 3.11 超时时正在执行哪一项。旧 selected-toolchain 工作流按现有条件 skipped，不计为成功 profile。

• W01 归属的第二次修复只改 V02 与 D05：延迟创建一个真实 bootstrap seed，为每个 case 完整复制独立 .git/工作树，并重新核对 HEAD/tree/clean。原 45 个测试函数及 210 个有序 case 保持，新增一项破坏单副本内容/ref/shallow/object 的隔离回归，seed 和另一副本完整字节保持。没有 guard、AST 或判定结果缓存。

• 干净本地检查点 d3940ead3805a903bbb711aa60f370225c81a9bb 完整 V02 为 211 passed，411.58 秒，零失败/错误/跳过，222 个源文件前后哈希一致。随后只更新 D05 已知证据；最终本地 0f5cf7b 与发布候选 7919288 同 tree。发布后真实 Git 历史、owner/stage、clean checkout、48 模块和 workflow 检查通过。

• 最终 Current verification run 38682280459, attempt 1 已 completed/success。Ubuntu 3.11 / 3.13、Windows 3.11 / 3.13 各实际 9,589 tests + 533 successful subtests，均零失败/错误/跳过。逐组 54 份原始文件、75 个测试文件、222 个前后相同源哈希、12 份完整零退出凭据、7 个实际 Git 提交、固定工具版本、6 份构建 inventory 和 W08/W14 安装见证均已独立核对。四组有序 collection 与源哈希完全一致；最终 API run/jobs/artifacts、PR head/base/draft 及未变 main 也完成聚合核对。

• 实际 pytest 耗时：Ubuntu 3.11 为 1,114.72 秒；Ubuntu 3.13 为 1,082.82 秒；Windows 3.11 为 1,963.36 秒；Windows 3.13 为 1,692.04 秒。现有 2,400 秒 suite / 50 分钟 job 预算保持。最终只读聚合 p4w02-final-hosted-review.json SHA-256 为 0c24ebeb9176d85e53dddef93e333afd58748cac36363afb53149cd23f8ab068。

范围与证据边界：

• 第一 W02 提交限于计划 14 条路径；验证修复单独使用已授权 W01 的 V02/D05；W02 后续仅 N01/D04/D05；这次 W01 初始化修复仅 V02/D05。批准计划 SHA-256 5fac7a2a796338faae9b98af6e0cc4519396e378ba10b8c5afeb4d029d97cd48 不变。没有改 guard、workflow、pins、预算、API、renderers 或后续产品模块。
• 首次失败的全部 192 原始文件、四个 ZIP/API digest、222 个源码哈希和精确测试身份已独立核对。W08/W14 个别安装见证存在且可协调，但最终 driver 因 earlier_stage_failed 未接受整次运行。
• 上传凭据不包含原始 GitHub event、依赖 wheels 或构建分发 archive 字节。W08 以完整成员等价核对，W14 可与 retained inventory 的完整 metadata 对齐，不声称 archive 字节重现。event 摘要在各 OS 内相同、跨 OS 不同，缺少原字节，未推断原因。
• 原 hero 保留其真实资源限制停止，有限私有案例完成不代表公共报告或发布完成。W03-W10、merge 与 release 尚未启动。

历史 W01 检查点：
f69cc5dd0a1de025dcf16ebdc4064cc1f20746b2 的 run 38536145728, attempt 1 四组全部成功；各 3,774 tests、另 533 successful subtests、73 文件、215 source hashes 和 12 份完整零退出凭据。先前 local 的 pip-check 原始退出记录不完整，原因未知，未追认其完整证明；单独复查和后续修复的 hosted 证据各按真实范围记录。被修复取代的旧 W01 run 38534506589 取消记录仍保留。基线仍为 PR #35 接受 merge fdad8c3449d2d223420c5d52bb9b4c1e863a9e3c。

phase4/implementation_evidence.json 只记录发布前已知事实。最终 exact-head hosted 结果记录在本 PR 及外部原始验收包中，没有追加自引用的“成功记录”提交。验收包保留两轮失败、全部最终原始工件、独立复核及逐文件 SHA-256 清单。