Lineage stage 2: trusted delivery persistence and semantic publication

[jeffhuber]

Part of #963, stabilization epic #979, and roadmap #900. Depends on accepted #990 (pure lineage contract). Replaces the delivery/publication portion of unaccepted draft PR #989.

Problem and outcome

Verified contribution history must survive takeover and subsequent fix rounds, be published through a trusted path, and be semantically read back before the active builder label moves. Missing or unreadable evidence must stop before side effects.

Scope

• Private metadata-only handoff/delivery persistence and current-head continuation recording, consuming Local builder handoff: make takeover quiescent, Git-write capable, and runner-consistent #962 quiescence/capability and Participants: enforce role-specific qualification for hosted builders #975 role qualification.
• Trusted publisher and builder auto-record attribution; selected explicit history remains distinct from REST numeric comment count and is validated before normalization.
• Semantic authenticated publication readback: parse the complete bounded trusted history, bind exact target/head and compare resolved episodes; substring presence or successful POST is insufficient.
• Publication-before-label reconciliation, fresh exact label/head reads, one active builder label, and no mutation on uncertain evidence.
• Delivery/publication functions and test harnesses required by maintained runners; actual rendered-runner tests must execute the failure paths without skip. Defer automatic runner label reconciliation activation and its maintained/template/package wiring to the consumer stage: accepted pre-lineage consumers must never observe moved labels under the old policy.
• Base-controlled provenance workflow, reviewed authority source/repository setting override, read-only permissions, no PR checkout/code/config execution, finite explicit comment-page requests and cap probe.

Keep provider reviewer admission, gate/status/labeler projections, the three automatic runner hooks, and CLI label-reconciliation activation for the consumer stage. New delivery/publication behavior may be exercised explicitly by tests but must not silently activate a label transition before all consumers are accepted. No new cloud schema or hosted implementation.

Typed-core dependency update

Consume the eventually accepted #990 contract from replacement PR #994. Superseded #989/#993 are regression references only. Keep canonical and standalone builder_lineage.py pure; storage, compatibility translation, authenticated publication and label application belong in narrowly owned producer adapters.

Construct explicit Target, Identity, History and Authorities values; feed public marker arrivals and private arrivals directly into one target-bound Chain factory before resolution. No optional-history fallback, pre-deduplication or intermediate budget reset. Missing transport refuses; genuine empty fetched history remains valid. Catch contract errors before any comment, label or artifact effect.

Translate supported legacy schema/writer-state records explicitly with full trust and immutable-target checks, or report unsupported history; do not silently discard them or weaken the pure parser. Writer-state strings are validated shapes, not proof of exit. Handoff requires separately verified #962 writer quiescence; logical completion/cancellation alone is insufficient. Suspended is not terminated, and continuation requires the actual same writer and chained exact heads.

Semantic publication readback uses authenticated public-only history and compares the complete canonical expected episodes. Adding private episodes to readback must not mask a missing publication. Keep automatic consumer/runner activation deferred to #992.

Acceptance and Code Mower delivery

One isolated Code Mower Codex writer, one draft child PR using Refs #963, accepted contract dependency on main, and qualified independent exact-head Claude review. Use the verified Codex runtime containment/Git/guard probe; the previously planned Claude source request was cancelled before launch, rather than adding a new runtime project to this unit. Codex execution has finite duration/output limits and no native USD cap. Reconstruct the owning preserved #989 diff, preserving relevant fixes and real CLI/generated-workflow/runner regressions. Cover takeover followed by ordinary fix, malformed transport/field/page shapes, empty/ambiguous/conflicting public markers, wrong target/head, readback trust, and refusal before comment/label/artifact side effects. Exact unittest discovery, focused checks, Ruff/privacy/package guards, full CI and authoritative gate must pass. Record writer exit/quiescence and lease closeout separately. No new hosted Devin create allowance.

Staged artifact tests must prove the new source behavior independently of the existing code-mower==1.4.0 template pin. #915 owns the eventual published-package pin and candidate/published installation qualification.

[jeffhuber]

Stage #991 is starting from accepted main e818a3b639dfe903bdc16aff3674af98a5a08233 with one fresh Code Mower Codex source writer and independent Claude review. The configured Codex model/settings are preserved. The previously proposed Claude source attempt was cancelled before launch.

Observed launch prerequisites passed: fresh held broker lease, eligible Codex builder admission, canonical Python runtime, dedicated Git write capability, protected pre-push/config boundary, and outside-checkout write denial. The initial source attempt is bounded to 1,800 seconds and 32 MiB output; no native USD cap is claimed and no retry or continuation is automatically authorized.

The scope remains 16 preferred / 18 maximum paths and 250,000 diff bytes: explicit trusted producer primitives and separately staged artifacts only. Accepted pure core/tests, default init/workflow/runner behavior, and live consumers remain unchanged. Pre-push guards enforce the envelope, protected core, and inactive integration boundary. Delivery is one draft PR closing #991 and referencing #963; root coordinates exact-head independent review and merge after deterministic checks, full unittest discovery, CI and gate.

[jeffhuber]

One bounded same-Codex completion continuation is authorized for the preserved #991 candidate. The initial source attempt exited after an existing shell fixture tried to use a temporary directory outside its permitted checkout; no commit, push, or PR was produced. Its exact dirty patch and untracked files were preserved, with observed writer quiescence and unchanged guards.

Trusted broker diagnostics passed 135 affected tests (one skip), 30 core tests, 20 state/publication tests, and the general guards. Three remaining artifact-test failures reduce to two assertions: use the actual normal-init emitted tools/builder_lineage.py helper for isolated import/producer-absence coverage, and compare exact parsed comment page values instead of matching page=10 inside per_page=100. Only that owning artifact test may change; production and other tests remain preserved.

The continuation is capped at 600 seconds/16 MiB under the same Codex permission profile and configured settings, with no native USD cap or automatic retry. If the corrected owning checks pass, the same author may commit the preserved candidate and open one draft PR. Trusted broker verification remains distinct from contained source execution. No full suite, independent review, activation, or merge is authorized by this continuation.