feat(alpha): first-run contract, flagship leak demo, and the retention-path witness MVP (arcs A1–A3) - #309
Conversation
… never a trace, never a clean scan
The alpha's worst first-run bug: an internal crash of the Python core
exited 1 (traceback), and `owen check` without --fail-on-finding mapped
rc<2 to 0 — a crash reported as a CLEAN scan. The whole failure surface
is now under contract:
Core (python -m ownlang):
- ownlang.__main__.run(): catch-all — internal crash exits 70
(EX_SOFTWARE) with one actionable line, no traceback; OWNLANG_DEBUG=1
prints the full traceback but KEEPS exit 70 (a re-raise would exit 1
and reopen the clean-scan hole). Deliberate contract errors keep their
codes. Documented in spec/CLI.md; pinned by tests/test_cli_contract.py.
- ownir.load(): an unreadable facts file is a polite OwnIRError (exit 2),
not an OSError traceback.
owen CLI:
- top-level catch (Program.cs) + CrashReport: internal error -> exit 5,
short message, deterministic diagnostic report at
~/.owen/diag/last-failure.json (tool/OS/runtime, command line, stage,
cause — NO source contents; facts sharing stays explicit --emit-facts);
--debug / OWEN_DEBUG=1 re-throws the full cause.
- core exit 70 -> framed politely -> exit 5 (the clean-scan hole closed).
- extractor output is captured: contract codes (0/2/4) pass through
verbatim; an unexpected exit code is framed as an internal error with
the raw trace in the report (or on stderr under --debug) -> exit 5.
- unknown options are a usage error (exit 2) with a hint, not phantom
paths ('owen check --verbose .' used to exit 4 'path does not exist').
- --emit-facts write failures are a polite exit 2.
- --help documents the full exit-code table incl. 5 and --debug.
Gate A (CI): new smoke steps — unknown option is exit 2; --help documents
the contract; a crash-injected core (OWEN_PYTHON wrapper, Linux) surfaces
as exit 5 politely and never as a clean scan, and --debug shows the full
cause while still exiting 5.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
The bug-report template asks exactly what triage needs (version, OS/SDK, command, exit code, output, optional diagnostic report / sanitized facts — with an explicit review-before-attaching note), and routes the documented exit-3/exit-4 behaviors away from the bug queue. The Known limitations section separates unsupported-by-design (frontend scope, symbol-tolerant analysis of non-compiling projects, alpha rule families, static-only evidence, versioned vocabulary, Python requirement) from bugs, per the alpha first-run contract. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
Benchmark publication (A1): scripts/benchmark.py gains --json — the publishable scorecard carrying the aggregate numbers WITH the corpus dirs/case counts, the git revision and the methodology (a metric without those is just a nice percentage nobody can audit). Under GitHub Actions the same numbers land in the job step summary; the corpus-benchmark CI job uploads the scorecard as a 90-day artifact. Shape and determinism pinned in --selftest. Flagship console repro (A2, examples/flagship/): the #278 shape reduced to ~80 cross-platform lines — a static publisher, a view subscribing in its ctor, an unsubscribe that EXISTS behind 'if (!keepAlive)' while every close path passes keepAlive: true. bad/ runs and prints '1000 still subscribed'; ok/ moves the release into Dispose() (provable teardown, mirroring the corpus-pinned predicate family) and prints 0. Both are dotnet-runnable and owen-checkable; gate A smoke-asserts bad → exit 1 + OWN001 and ok → exit 0 against the installed tool on Linux and Windows. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
The published contract says every internal failure exits 5, but the top-level handler re-threw under --debug/OWEN_DEBUG=1 — the process then died with a runtime-chosen, platform-dependent code. Debug now changes the VOLUME of diagnostics (the full exception is printed to stderr), not the machine semantics: report written, exit 5, both modes. Gate A gains the previously-untested top-level-exception smoke (TMPDIR pointed at a missing directory fails Path.GetTempFileName before any inner catch): polite + no stack by default, full stack under --debug, exit 5 in both. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
…inus the dominator tree The runtime-witness half of the flagship demo, resurrected as the minimal core rather than a rebase of the geological layer: Heap.cs (mark-from-roots census, sampled ranked retainers with the field named at every hop, stack/finalizer vs real retention) + Program.cs (census / roots, --pid|--dump, --type, --sample/--max-hops limits, --out runtime.json) + the csproj. Deliberately NOT taken: Dominators.cs (444 lines of separate proof surface the hero demo does not need), the pre-#278/#305/#306 corpus slice, and the old runtime-contract docs. Adapted to the A3 contract: - verdict vocabulary — RETAINED (a root path was established, exit 1) | OBSERVED_ONLY (instances exist, no path established — never called a proven leak, exit 0) | ABSENT; usage documents it; - the JSON artifact carries verdict + collector identity (mode attach/dump, target, runtime, OS — no timestamps: identical heaps must yield identical artifacts); - a failed read stays exit 2, distinct from clean(0)/retained(1); - net8.0 cross-platform (ClrMD 3.x live-attach works on Linux and Windows; legacy net472 heaps are dump-only from here — the PR #280 live-attach build can be resurrected the day a live legacy target matters). Gate A builds the project and smokes the honest-failure surface on both platforms: bare usage exits 2 with the verdict vocabulary documented, and a failed attach exits 2 — never a clean scan. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
…ry; live static-event verified The first live net8 run exposed two masked layers, both now fixed and pinned: 1. Root seeding masked the durable evidence. The static publisher is BOTH a stack-root object (a Main local holds it in a register) and reachable from the pinned statics object[]; single-pass seeding claimed it as a [stack] root before BFS could reach it durably, so the flagship's static-event retention read as 'live right now, not retention' and the dominant-retainer callout was suppressed. FindRetainers now runs TWO BFS PHASES: durable roots (handles/statics) are seeded and walked to exhaustion before any transient root (stack/finalizer) enters the graph — retention analysis prefers durable evidence; the stack only explains what nothing durable can. 2. The classifier is now an explicit boundary (Retainer.Classify), pure over (root kind, path evidence): every known ClrRootKind is named; delegate evidence (EventHandler/MulticastDelegate hop types, the _invocationList FIELD — types checked as types, fields as fields, so a type merely named '...StackMachine...' is not evidence) upgrades pinned/handle roots to static-event; an UNKNOWN kind reports 'unsupported-root:<kind>' — honest refusal, never silently non-root. Pinned at two levels: - 'RetentionPath selftest' — 7 heap-free classifier fixtures: the verbatim live net8 static-event path, the stack-flavoured-names negative neighbour, stack/finalizer doctrine (delegate evidence does NOT upgrade a genuine stack root), plain static-field, and the honest-refusal case. Gate A runs it on both platforms. - Gate A end-to-end (Linux): the flagship bad app held live (OWEN_FLAGSHIP_HOLD=1 — new opt-in hook on both samples, also what the demo orchestrator will use), retention-path attaches and must report verdict RETAINED via static-event with exit 1, in the human output and the JSON artifact both. Verified live locally before CI: the same attach that read '[stack], 5 hops / no single dominant retainer' now reads '[static-event], 6 hops (object[] -> AppSettings -> PropertyChanged -> _invocationList -> handler -> DocumentView) / 100.0% hang off ONE reference'. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
scripts/flagship-demo.sh — the reproducible demo entrypoint, kept a DUMB orchestrator per the arc contract: build sample+witness, hold the app live (OWEN_FLAGSHIP_HOLD), attach retention-path through its PUBLIC CLI with a timeout, machine-validate the JSON artifact (python, not grep) against the human verdict — a disagreement is a failure in itself — assert stable semantic anchors (AppSettings / PropertyChanged / _invocationList / DocumentView / static-event), never verbatim paths or addresses, print ONE stable summary line, clean up via trap. No classification logic of its own. Running the ok variant through it immediately caught a real verdict gap: the last loop iteration's view stays live in Main's frame, and the witness reported RETAINED for a [stack]-only path — the verdict did not consult the classification. Per the transient-root doctrine a stack path proves 'live right now', not retention: 'roots' now requires at least one DURABLE retainer for RETAINED/exit 1; all-transient reachability is OBSERVED_ONLY/exit 0 with the transient paths shown as explanation. 'unsupported-root:*' counts as durable on purpose — unknown evidence surfaces loudly, never quietly demotes the verdict. Also pinned in code: the two-phase-BFS invariant (durable traversal claims reachable targets; transient traversal may share intermediates but cannot overwrite durable ownership — with the exhaustion proof). Gate A (Linux) now runs the orchestrator itself for both variants — bad: 'leak DEMONSTRATED: static-event retention'; ok: 'fix VERIFIED: no established retention' — so the demo entrypoint and the CI proof are the same artifact. Both verified live locally before push. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
The orchestrator failed on the runner exactly at the predicted risk point: ubuntu runners ship kernel.yama.ptrace_scope=1, which blocks same-user non-ancestor PTRACE_ATTACH, so ClrMD could not attach to the held flagship app (locally the session runs as root, where Yama does not apply). The step now sets scope 0 first — a CI-runner-only relaxation; the demo script itself stays sudo-free, and a real user under scope 1 gets the witness's polite exit 2, never silence. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
…nned; ok-demo checks the user contract
Arbiter refinements on the A3/demo vertical, all three:
1. The JSON artifact is no longer RETAINED-only: ABSENT and
OBSERVED_ONLY write the same runtime.json (verdict + collector
identity + any transient paths) when --out is given — the ok side of
a demo is now as machine-checkable as the leak side, via one shared
WriteArtifact writer.
2. The verdict rule is a pure function (IsDurableKind / VerdictOf) and
the selftest pins it deterministically: stack-only and
finalizer-only reachability are OBSERVED_ONLY; one durable retainer
makes RETAINED; unsupported-root:* surfaces as RETAINED, never a
quiet demotion. 11 heap-free checks now.
3. The ok-demo validator asserts the USER contract only — exit 0,
JSON verdict in {ABSENT, OBSERVED_ONLY}, zero durable retainers,
human/JSON agreement — instead of insisting on OBSERVED_ONLY: which
of the two shows up depends on JIT liveness of a loop local, an
internal CLR decision that is not a public API. The verdict
SEMANTICS stay proven by the deterministic selftest; the live demo
proves the absence of a false RETAINED. Also: trap extended to
INT/TERM.
Verified live locally: selftest 11/11, demo bad 'leak DEMONSTRATED:
static-event retention', demo ok 'fix VERIFIED: no established
retention (OBSERVED_ONLY)'.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
|
Warning Review limit reached
Next review available in: 8 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (20)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3774e5f6c2
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| foreach (var o in Heap.EnumerateObjects()) | ||
| { | ||
| if (!o.IsValid || o.Type?.Name == null) continue; | ||
| if (!IsType(o.Type.Name, typeName)) continue; | ||
| totalOfType++; | ||
| if (targets.Count < sample) targets[o.Address] = o.Type.Name; |
There was a problem hiding this comment.
Selecting the first sample objects in heap-enumeration order can produce a false OBSERVED_ONLY verdict. For example, if 200 older instances are unreachable garbage and a later instance is held by a static, the default sample contains only garbage, SampledRetained becomes zero, and the actual durable retention is never examined. Use an unbiased sample spanning all matching instances, or continue until enough reachable instances have been collected.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed and fixed in 2f064dd: ALL matching instances are targeted now — the verdict is exact over the whole population; only path resolution stays sampled (bounded at --sample paths). A later durably-held instance can no longer hide behind older garbage.
Generated by Claude Code
There was a problem hiding this comment.
Follow-up in f389b92 — the 2f064dd fix made reachability exact but left durable/transient classification on the resolved-path sample, which re-admitted the same bias one level up (200+ finalizer-queue-reachable corpses ahead of one durably-held instance in heap order → false OBSERVED_ONLY). Now a root-kind census walks the parent chain for EVERY reachable instance (dictionary hops only, no name resolution) and the verdict consults the exact durable count; path resolution stays bounded by --sample, durable instances resolved first so RETAINED always ships with a durable root path. Rendered shares now divide by resolved paths, not the population. Selftest 11 → 16: the ClrMD-level transient split and the string-level verdict rule are pinned to agree on every root kind.
Generated by Claude Code
| string signature = string.Join(" -> ", hops.Select(h => h.Type)); | ||
|
|
||
| if (!groups.TryGetValue(signature, out var retainer)) | ||
| { | ||
| retainer = new Retainer(hops, kind); |
There was a problem hiding this comment.
The grouping key contains only hop type names, while the retained Retainer keeps the root kind of whichever matching path is encountered first. If stack-rooted and durably rooted instances have the same type sequence, they are merged and later verdict logic classifies the whole group using that first root; this can either hide a real durable retainer or turn transient reachability into a leak. Include the root kind—and ideally the distinguishing fields—in the signature.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed and fixed in 2f064dd: the group signature now carries Retainer.Classify(kind, hops) plus the per-hop fields (Hop.ToString), so stack-rooted and durably-rooted paths sharing a type sequence can no longer merge and inherit the first path's classification.
Generated by Claude Code
| if (!parent.TryGetValue(cur, out ulong p) || p == 0) break; | ||
| cur = p; | ||
| if (chain.Count > maxHops) break; | ||
| } | ||
| kind = rootKind.TryGetValue(cur, out var k) ? k : ClrRootKind.None; |
There was a problem hiding this comment.
When a path exceeds maxHops, unwinding stops at an intermediate object, so rootKind lookup yields ClrRootKind.None. Classify then turns that into unsupported-root:None, which IsDurableKind treats as durable; consequently, a target reachable only from a stack path longer than the limit is incorrectly reported as RETAINED. Path truncation must not discard the root classification used for the verdict.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed and fixed in 2f064dd: on truncation the (acyclic) parent chain is now walked to the true root WITHOUT recording hops — the rendered path stays bounded at --max-hops, but the verdict sees the real root kind. A long stack-only path can no longer surface as unsupported-root:None → durable → false RETAINED.
Generated by Claude Code
| @@ -0,0 +1,34 @@ | |||
| <Project Sdk="Microsoft.NET.Sdk"> | |||
There was a problem hiding this comment.
This adds a new audit collector and makes it an actively built and executed CI component inside audit/, even though the repository explicitly designates the separate OwnAudit repository as the home and source of truth for active audit development until consolidation. Implementing the witness here creates the divergent source-of-truth state that the repository rule is intended to prevent; move this work to OwnAudit or perform the documented consolidation first.
AGENTS.md reference: AGENTS.md:L17-L17
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Placement was deliberate, not a drive-by, so answering rather than moving: (1) the decoupling half of the AGENTS.md rule IS honored — RetentionPath has zero coupling to ownlang and is consumed only through its public CLI/JSON; (2) the witness core originates from this repo's own PR #280, which placed it in audit/runtime/ — extracting its minimal core onto the current head (explicitly directed during the A3 arc) continues that lineage rather than forking OwnAudit; (3) the flagship demo and its gate-A end-to-end smoke need the witness buildable next to examples/flagship/, which OwnAudit cannot provide without cross-repo CI. Whether the consolidation endpoint is "witness moves to OwnAudit" or "audit development consolidates here" is an owner decision the AGENTS.md note itself defers ("currently lives") — flagging it for that decision rather than relocating unilaterally.
Generated by Claude Code
| Console.Error.WriteLine( | ||
| " Re-run with --debug (or OWEN_DEBUG=1) for the full traceback, and " + | ||
| "please report it: https://github.com/PhysShell/Own.NET/issues/new/choose"); | ||
| return CrashReport.ExitCode; |
There was a problem hiding this comment.
When the Python core catches an internal exception and exits 70, this branch returns exit 5 directly without invoking CrashReport.Child or otherwise writing ~/.owen/diag/last-failure.json. Thus the most important internal-error path contradicts the new help text and failure contract, and users may be directed to submit a report that is absent or stale from an earlier failure.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed and fixed in 2f064dd: the rc-70 branch now goes through CrashReport.Child("analysis core", …), so the diagnostic report at ~/.owen/diag/last-failure.json is written on the most important internal-error path too. The gate-A crash smoke additionally asserts the report is mentioned.
Generated by Claude Code
…gnature, truncation root kind, core-crash report Four verified findings from the #309 review, all real: 1. Sampling bias (P1): taking the first `sample` instances in heap-enumeration order could catch only old garbage and miss a later durably-held instance — a false OBSERVED_ONLY. ALL matching instances are targeted now: the verdict is exact over the whole population, and only PATH RESOLUTION is sampled (bounded at `sample` paths). 2. Group signature (P1): grouping by hop type names merged stack-rooted and durably-rooted instances sharing a type sequence, classifying the whole group by whichever path came first. The signature now carries the classification and the per-hop fields. 3. Truncation (P1): stopping the unwind at maxHops yielded ClrRootKind.None -> 'unsupported-root:None', which the verdict counts as durable — a long stack-only path read as RETAINED. The parent chain is now walked to the true root for the verdict; only the RENDERED path stays bounded. 4. Core-crash report (P2): the rc-70 branch returned exit 5 without writing ~/.owen/diag/last-failure.json, contradicting the published contract on the most important internal-error path. It now goes through CrashReport.Child (null capture — the core's own polite line is already on stderr), and the gate-A crash smoke asserts the report is mentioned. The fifth finding (AGENTS.md: audit development lives in OwnAudit) is a placement decision answered on the thread: the decoupling half of the rule is honored (no ownlang coupling, public CLI/JSON only), the witness core originates from PR #280 which placed it in this repo's audit/, and the extraction here followed an explicit owner-side direction; whether to relocate to OwnAudit stays an owner decision, not a drive-by move. Verified locally: builds clean, selftest 11/11, flagship-demo bad 'leak DEMONSTRATED' / ok 'fix VERIFIED (OBSERVED_ONLY)'. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
…o longer flip the verdict Completing the Codex P1 sampling fix one level up. After 2f064dd the REACHABILITY count was exact, but durable/transient classification still came only from the resolved-path sample (first --sample reachable instances in heap order). A type with 200+ finalizer-queue-reachable corpses sitting ahead of one durably-held instance would still read as a false OBSERVED_ONLY — the same bias, moved from discovery to classification. Now FindRetainers runs a root-kind census over EVERY reachable instance (parent-chain walks only — dictionary hops, no type/field resolution), and the verdict consults the exact DurableRetained count. Path resolution stays bounded by --sample, resolving durably-rooted instances first, so RETAINED never ships without a durable root path on display. --sample/--max-hops are clamped to >= 1: a pathological `--sample 0` must not suppress the path a RETAINED verdict promises. Display honesty fixed with it: rendered shares now divide by the resolved-path count, not the full population (200 resolved paths out of 1,000 reachable used to render a 100%-dominant retainer as 20%), and the human output states all three numbers separately — on heap, reachable, durably retained, paths resolved. The invariant is pinned in code at the API boundary: sampling and hop limits affect evidence PRESENTATION, never discovery, classification, aggregation, or the exit code. Selftest grows 11 -> 16 checks: the ClrMD-level transient split (BFS seeding + census) and the string-level verdict rule must classify every root kind consistently, unknown kinds included. Verified locally: build clean, selftest 16/16, flagship-demo bad (RETAINED via static-event) and ok (OBSERVED_ONLY) both pass; manual attach with --sample 50 against 1,000 leaked views reports "1,000 durably retained (50 paths resolved for display)". Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015de4MezSeUnZBoWq1fFU5M
|
@coderabbitai review |
2 participants
Что и зачем
Подготовительный пакет Owen Alpha — три закрытые арки, каждая с исполняемым доказательством:
A1 — контракт первого запуска. Закрыта настоящая soundness-дыра оболочки: краш Python-ядра давал exit 1 (traceback), и
owen checkбез--fail-on-findingмаппил его в 0 — clean scan. Теперь: ядро — catch-all → exit 70 (EX_SOFTWARE), одна вежливая строка (OWNLANG_DEBUG=1печатает traceback, но код остаётся 70); CLI — top-level catch → exit 5, diagnostic-отчёт в~/.owen/diag/last-failure.json(без исходников; facts — только явным--emit-facts);--debugменяет объём диагностики, не машинную семантику (exit 5 в обоих режимах). Вывод extractor'а буферизуется — контрактные коды проходят как раньше, неожиданный код фреймится вежливо. Опечатки флагов — usage error,--helpдокументирует полную таблицу кодов. Плюс: EN issue-шаблон «owen CLI problem», секция Known limitations (unsupported-by-design ≠ баг), публикуемый benchmark-scorecard (--json+ step summary: числа с корпусом, ревизией и методикой).A2 — flagship-репро. Сюжет #278 в ~80 кроссплатформенных строках (
examples/flagship/console/): статический publisher, подписка в конструкторе,-=существует заif (!keepAlive)— а каждый Close передаётtrue.bad/сам печатает «1000 still subscribed»,ok/(release вDispose()) — «0». Обаdotnet run-able иowen check-able; opt-in hold-хукOWEN_FLAGSHIP_HOLD=1для witness/CI.A3 — witness MVP.
audit/runtime/RetentionPathизвлечён из PR #280 (минус dominator tree и устаревший corpus-срез), ретаргет net472→net8.0 (Linux/Windows live-attach через ClrMD). Два содержательных фикса, найденных живыми прогонами: (1) двухфазный BFS — durable-корни (handles/statics) обходятся до исчерпания раньше transient (stack/finalizer), иначе локаль Main в регистре присваивала статический publisher как[stack]и маскировала static-event retention (инвариант с доказательством зафиксирован в коде); (2) вердикт консультируется с классификацией —RETAINED/exit 1 требует durable-retainer'а, чисто transient-достижимость —OBSERVED_ONLY/exit 0 (loop-local — не утечка). Классификатор — явная границаRetainer.Classify: известныеClrRootKindперечислены, delegate-evidence по структуре пути (тип «...StackMachine...» — не evidence), неизвестный kind →unsupported-root:— честный отказ. JSON-артефакт (verdict+ collector identity) пишется для каждого вердикта.Пост-ревью (Codex + follow-up).
2f064dd: таргетируются ВСЕ инстансы типа (не первые N в порядке кучи), сигнатура группы путей несёт классификацию корня и поля хопов, обрезка--max-hopsдоматывает parent-chain до истинного корня, rc-70 ядра пишет diagnostic-отчёт через общийCrashReport.f389b92(завершение инварианта): вердикт берётся из точного root-kind-ценза по каждому достижимому инстансу —--sample/--max-hopsограничивают только отображение и математически не могут изменить вердикт/exit code; durable-инстансы разрешаются первыми (RETAINED всегда с durable-путём на экране); доли рендерятся от разрешённых путей.Demo-оркестратор
scripts/flagship-demo.sh— тупой по контракту: build → hold → witness через публичный CLI с timeout → машинная JSON-валидация против human-вердикта (расхождение — fail) → семантические якоря, не дословные пути → одна стабильная итоговая строка → trap-cleanup. Gate A запускает его для обоих вариантов на Linux (с CI-only послаблением Yama ptrace_scope; сам скрипт sudo-free — attach, запрещённый политикой ОС, остаётся честным exit 2).Тип изменения
Как проверено
python tests/run_tests.py— все сьюты, включая новыйtests/test_cli_contract.py(4 проверки контракта ядра)ruff check .иmypyscripts/benchmark.py --selftestрасширен (scorecard, 17 проверок) и гоняется в lint-джобе;RetentionPath selftest(16 фикстур: классификатор, вердикт, когерентность ClrMD- и строкового уровней доктрины) гоняется в gate A на обеих платформахRETAINED [static-event], 6 hops; с--sample 50против 1000 вью — «1,000 durably retained (50 paths resolved for display)»; ok →OBSERVED_ONLY), оба варианта оркестратора зелёныеf389b92)Связанные issue
Refs #250/#252/#253/#254 (Owen Alpha: этот пакет закрывает инженерную часть подготовки; владельческий гейт #252 — лицензия/версия/environments/секрет — остаётся единственным блокером публикации), #278 (сюжет flagship-репро), #280 (источник witness-ядра; dominator tree и старый corpus-срез сознательно не взяты), #270 (runtime witnesses).
Чеклист
feat:,fix:,docs:…)Generated by Claude Code