镜像站点 · 本页由第三方 GitHub 只读镜像提供,非 GitHub 官方站点,不接受任何登录或凭据输入。前往 github.com
Skip to content

fix(ci): run unit shard diagnostics through supported owned tooling #1131

Closed

Description

@chrisbbreuer

The exact-source CI 43254917700 at 0e6e4f9250bd9b2af71968d0391bd20c3593ff3a finishes with 56 successful jobs, the unchanged TSan unit shard 1 deadline, and a separate dependency-audit failure. Audit job 131316233728 rejects tools/unit78-native-diagnostic.py: uses an unclassified runtime extension.

The earlier pinned-package error messages in this job belong to expected negative controls. The actual repository failure is the newly introduced observer runtime. docs/.data/dependency-inventory-v1.json permits the owned Home TypeScript script runtime; the completed tooling migration (#497) does not permit this Python helper. This was missed in the diagnostic source qualification.

Acceptance:

  • Replace the observer with a supported owned tool implementation. Use the existing Home TypeScript APIs where sufficient, or the repository's native Zig platform-tool pattern for process/proc/debugger operations. Improve the owning runtime if a required primitive is absent; do not add shell/Python substitutes, broad runtime exceptions, or weaken the audit.
  • Preserve the exact pinned engine source, original full shard command/fixture, both JIT selections and all 5,001 calls, allocator/heap/ballast checks, TSan instrumentation and original deadlines. Observation remains separate from qualification.
  • Preserve process identity checks, bounded proc samples/GDB captures, helper cleanup/detach, capped credential-free evidence, early/terminal retention and propagation of the original child status. Retain original artifacts and failure evidence.
  • Cover actual owned harmless children and timeout/reaping/identity/error cases, register the implementation and callers in the existing dependency/tool contracts, then pass the actual unchanged dependency-audit before normal source delivery.

Related #1129 (native unit shard completion), #462 (dependency boundary), #497 (owned tool migration). The original diagnostic helper and its GDB/setup commits remain preserved history. This issue fixes the repository-tooling failure and does not claim a native engine performance fix. Retained job log: Code/Verification/settlers-retained-replay/native-growth-audit/managed-frames-integration/tsan-unit-completion/latest-main-dependency-audit-job.log.

Activity

  1. chrisbbreuer commented on Nov 13, 2026

    @chrisbbreuer
    MemberAuthor

    The supported native observer landed in 96bdc76. Local canonical audit passed its 7 unit tests and 9 portable self-test groups; the Linux target was compile-checked only.

    The exact commit's pending CI run was naturally superseded before starting jobs. Its immediate descendant 4cc4e29 changes only ecosystem workflow pins; all seven native diagnostic source files and the audit inputs are unchanged. That descendant's actual Linux dependency-audit job failed with ResistantFixtureNotReady at the unchanged five-second resistant-child lock readiness check. The boundary audit itself classified 16 edges and 65 tools successfully, and all 7 unit tests passed. This is not a complete Linux self-test pass.

    The raw 46,163-byte job log is retained with SHA-256 a357ae9d400514d2c256d67cbaa17b24e8b826d110435109ee99cce9371a812e. It does not identify the failing policy, child startup error or identity publication state. The next source change will expose bounded failure-path evidence while preserving the original readiness deadline, closed-pipe cases and process cleanup rules. No CI retry or deadline increase has been made. This issue and #1129 remain open.

  2. chrisbbreuer commented on Nov 13, 2026

    @chrisbbreuer
    MemberAuthor

    Delivered efdcad8656. The native helper now retains bounded, failure-only startup evidence without changing the original readiness threshold, process controller, engine pin or evidence limits. The startup reader has real file/schema/ownership controls; diagnostic group labels never authorize signals.

    The same delivery includes the omitted canonical documentation-manifest refresh. The exact pinned BunPress 0.2.14 renderer produced the three expected changed outputs, matching the original hosted hashes, with the other 52 unchanged. A checked build verified all 55 physical outputs and the complete tree digest. The native docs-link gate passed 5 build steps and 4 parser tests; a separate invocation of its exact emitted checker confirmed 378 files with all links resolving. Original verifier failures remain retained separately from successful native operations; generation and compilation were not retried to replace them.

    The unchanged helper bytes retain both independent review approvals and the prior local native evidence: 9 build steps, 10 unit tests and 9 portable process-group fixtures passed. Linux compilation was checked separately; those host results do not establish Linux execution or explain the original ResistantFixtureNotReady incident.

    Actual new-source CI 43368539774 is pending behind the existing main queue. Ecosystem 43368542045 is running; its final integration/package evidence is not yet accepted. This issue remains open until required Linux audit evidence is available. #1129 and the unchanged full TSan shard remain separate outstanding work. No original workflow rerun or cancellation was requested.

  3. chrisbbreuer commented on Nov 13, 2026

    @chrisbbreuer
    MemberAuthor

    The exact efdcad8656d828b5fd8abe6486e45b534677983e Linux dependency audit failed again. The new failure-only evidence narrows the boundary: policy term-open, startup stage identity_publish, error MalformedProcStat. The producer's group matches the owned leader. The captured stack reaches parseStat at the missing-closing-parenthesis check through statRecord, processStamp and writeFixtureStamp. The identity file was not published; the original readiness deadline remains unchanged.

    The final summary is 7/9 build steps and 10/10 registered tests passed, with the diagnostic self-test exiting 1 as ResistantFixtureNotReady. This is not an audit pass. The separate docs job passed its 378-file link check and manifest-checked 51-page build. Raw audit log SHA256: ad449c8f40adfe4f559ccd9778079c25599bdc4f1077e56e8a29b7fd67828a8a.

    Source investigation identifies a relevant pinned API behavior: the allocating writer can return EOF from cached file size 0 before reading content. Linux procfs can generate readable content independently of that size. The historical proc bytes and metadata were not captured, so empty input in this particular run is not directly established.

    A source repair and real Linux controls are being prepared for all five proc readers using bounded streaming, preserving strict parsing, existing caps, PID/start/group checks and lifetimes. A separate hash-reader EOF handling defect is also being tested explicitly; it is not claimed as this observed startup cause. No source repair, new Linux pass or completion is claimed yet. The original run has not been retried or cancelled.

  4. chrisbbreuer commented on Nov 13, 2026

    @chrisbbreuer
    MemberAuthor

    Delivered in 8ec6bfa8, through #1132.

    The unchanged zig build dependency-audit passed on the reviewed candidate before main delivery in PR audit job 131751567497, then passed again on the delivered main commit in main audit job 131757441321. Main reports 16 classified edges, zero migration targets, 65 classified repository tools, and all 9 portable plus 3 Linux observer lifetime fixtures with zero exclusions. The main docs job also passed.

    All five virtual proc readers now use bounded streaming without treating reported file size as EOF. File hashing handles the pinned streaming API’s EndOfStream. Strict parsing, PID/start/group identity, read caps, deadlines, cleanup and original shard commands remain intact. Real empty/abc digest failures and file-cap/error controls are retained; the local complete diagnostic units passed 13 cases with 3 Linux-only skips, followed by the actual Linux audit above.

    Closing the repository-tooling failure. The broader CI and ecosystem runs are still in progress, and #1129 remains the separate native shard completion investigation. This does not claim an engine performance or game compatibility fix.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions