Skip to content

[v0.5] Adopt PrismPM's Prism factory for development, CI and release tooling #468

Open

Description

@afflom

LexLean v0.5 must use PrismPM's Prism model and factory to reduce development, CI and release overhead. This is an explicit release requirement from the project owner, linked to #128. Installation alone or another wrapper workflow does not satisfy it.

Pinned starting point

PrismPM v0.4.0 is published at b48ee2ffa4d2097685054dc350b57aa204fc7080. Its Factory.Release model and factory adapter describe source/build routing and establish or reuse original successful ancestor evidence when the relevant input identity matches and that evidence remains available. Changed release plumbing is checked separately. The adapter is repository-specific, not a generic command that can be dropped into LexLean.

The broader generated workflow/scheduler in PrismPM#586 has merged as 8ccb0f96d30928c4083f1970094149ee5af3b3c4. It is now an available source reference for adaptation. This merge does not establish that it shipped in the pinned v0.4.0 release, that its evidence is qualified for LexLean, or that its scheduler and reuse policy are admitted here. Preserve the separately reviewed LexLean model, correspondence, full-acceptance and ownership boundaries below. The published SDK vendors LexLean e0b74b81e6f49c83596a957d232c8383a5db1a25; publication does not establish delivery of #466 into that SDK or close #465/#454.

Delivery slices

  • Model and ownership. Adapt the published Prism factory model to LexLean's actual developer, producer, CI and release obligations. Inventory each obligation's inputs, outputs and sole executing owner. Prove the routing/invalidation properties under explicit modeled assumptions and test the adapter's correspondence to that model. Qualify external evidence admission and workflow execution separately; finite model proofs alone do not certify GitHub or shell execution. The model must control real routing; a descriptive model alongside independent YAML is insufficient. Identify and retire the orchestration it replaces.
  • Developer/producer integration. Coordinate the supported developer entry point with ci: regenerate.yml, the write recipes on runners #467's existing owner. Use its official write recipes and reviewable base-bound output handoff, with explicit Atlas scope, complete writer results and shared-path rejection. Preserve one admitted identity wave; do not run a second set of writers. Record partial/failed output honestly.
  • CI and release evidence reuse. Adapt the factory's identity/receipt boundary to LexLean rather than copying PrismPM's job names/counts. Bind source tree, relevant compiler/model/toolchain/profile/platform inputs, producer/check definitions, and upstream identities. Reuse only original successful evidence whose identity and retained artifacts satisfy the contract; never synthesize success from canceled, skipped, expired or partial runs. Fresh model/plumbing verification owns changes excluded from other identities. Preserve the required independent clean reproducibility builds. Publication consumes qualified immutable products instead of rebuilding identical products merely because the workflow entered another stage.
  • Qualified v0.5 integration and measured removal of overhead. Land reviewed implementation through the actual release/forward history. Demonstrate unchanged-input reuse and correct invalidation for semantic/toolchain/producer and plumbing changes. Use existing receipts as the baseline where applicable; report actual wall time, executed work, setup steps and storage costs with comparable scope. Remove the demonstrated duplicate work and obsolete workflow path. Document the supported local/hosted entry point and recovery from missing evidence.

Acceptance that must remain intact

Every existing proof, axiom-policy, conformance, host/Atlas, package and release obligation retains an explicit owner and evidence. Current full final-head acceptance, same-run honesty and independent reproducibility remain required; no old-head receipt automatically substitutes for them. Any proposed normative gate-reuse change needs a separately reviewed specification/model/correspondence change establishing equally strong acceptance before rollout.

Test the new routing boundary with mismatched identity, unavailable evidence, failed/canceled producers and shared-path/partial-output cases. These are focused factory tests, not permission to repeat completed LexLean producers or gates. Compare the reviewed, tested and actual merged trees and retain immutable provenance. Scheduling heuristics must never decide acceptance.

Coordination and release completion

Root owns this new integration/model/evidence-contract scope. Claude retains #467 implementation/qualification, #405 forward integration and existing contributor work; shared GitHub authorship is not takeover permission. Coordinate concrete file and execution ownership before edits that overlap those lanes. Reuse delivered Intel construction/transport and existing #459/#462 work rather than reimplementing them. Preserve #316's frozen evidence and its separate forward-integration dependency.

This issue blocks #128 until the Prism-driven tooling is actually used for LexLean's supported development, CI and release path on release/v0.5, redundant orchestration is retired, and the resulting implementation is qualified. An initial regeneration pilot is a delivery slice, not closure of this whole issue. #128's existing #106/#15 and §30 release criteria remain mandatory. No numeric speedup is promised before measurement.

Activity

  1. added this to the LexLean v0.5 milestone on Nov 11, 2026
  2. afflom commented on Nov 11, 2026

    @afflom
    MemberAuthor

    Source-preparation progress: the pinned release/v0.5 inventory now maps 62 distinct whole-check, producer and release obligations to inputs, outputs and existing owners. The initial RP-19 register/scenario/named-test design uses the authoritative spec/src inputs; generated SPEC/model files have not been hand edited or projected.

    Review corrected an unnecessary-rerun rule: acceptance obligations delegate to the existing gate owner and its exact-current-head/same-run/repro receipts, rather than automatically establishing a completed gate again. No old-head receipt becomes acceptance, and routing a missing obligation is not admission to launch a producer.

    A second review found the proposed Bash-only boundary unsuitable for portable policy coverage. Host dispatch is explicitly unresolved before adoption: portable model/policy assertions must cover supported hosts, while any genuinely POSIX adapter checks use the existing documented reporting contract. No blanket skip was added. The stale source-preparation generator was corrected without running it; superseded candidates/reviews are preserved.

    Next bounded work is that portable policy/adapter split, followed by review of the official projection and first named-failure execution proposal. No projection, compilation, runtime test, identity writer or implementation has been launched. This remains initial source preparation, not completed factory integration. #467/#405 owners and #316's preserved evidence remain untouched.

  3. afflom commented on Nov 11, 2026

    @afflom
    MemberAuthor

    Portable source review found that the proposed sibling xtask executable is not guaranteed on every supported CI host; those jobs intentionally avoid that separate build. I am not adding an all-host build merely to satisfy the new test. The existing linked compiler evaluator already provides the production-lowering/interpreter path used by portable stdlib conformance, so that is the intended policy-test boundary. xtask remains a thin separately qualified adapter.

    For the initial R3 observation, the reviewed local native-boundary test remains usable only with an authenticated candidate executable, official register projection and matching named test. Its intended failure is the real unsupported factory-route capability, not a missing artifact or setup error. Independent review confirmed that local development red need not itself establish final all-host qualification. The same claim and 17 expectations must survive the later split: portable model assertions plus retained native-adapter correspondence coverage must both turn green. The initial test will not be discarded or weakened.

    The published Prism model lacks the richer status/artifact/origin/obligation inputs, so I will not manufacture a semantic counterexample by passing inadmissible records as already-valid receipts. Source-only preparation now covers the concrete local R3 lifecycle and preservation/resource requirements. No projection, build, test, policy implementation, cleanup or writer has been admitted. #467/#405 remain with their owners, and all #316 evidence remains untouched. Final model proof/correspondence, real developer/CI/release consumption and measured retirement of duplicate orchestration are still required for closure.

  4. afflom commented on Nov 11, 2026

    @afflom
    MemberAuthor

    The next source plan now provisions the matching native formatter and xtask together in one candidate checkout and one official Cargo build. The preserved historical formatter cannot be reused: its embedded semantics differ from the committed v0.5 spec lock, which fmt validates. Building a separate formatter checkout would add avoidable compilation and provenance work.

    Independent review confirmed the combined build preserves register → scenario → named failure → implementation ordering. The revised driver is source-only preparation: capture both ordinary native executables from the same Cargo JSON, stop for exact artifact/source/feature/profile authentication, then separately admit formatting/projection. This avoids a duplicate probe build merely to discover Cargo-hashed output filenames. No factory implementation or execution has started; all 17 expectations and the portable linked-evaluator/native-adapter follow-through remain intact.

    Capacity is still unresolved. Current available space was639,414,272B. A bounded review found one conditional completed-checkout sparse-conversion candidate with at most475,512,832B reclaim; even that optimistic sum is54,705,744B short of the source/output planning scenario, and983,560,784B short of the scenario including final frozen tools and illustrative link scratch. These are planning estimates, not measured peaks or hard limits. Additional inspected roots lacked exact completed-carrier evidence and were excluded. No conversion or cleanup is admitted just to make a partial dent in an unfunded phase.

    #405/#467 ownership, all #316 evidence and stopped work remain preserved. Matching-tool provisioning, actual resource/control checks and a separately reviewed phase admission remain prerequisites. Full checked-model/proof/correspondence, real developer/CI/release consumption, measured removal of duplicate work and final qualification remain required for closure.

  5. afflom commented on Nov 11, 2026

    @afflom
    MemberAuthor

    The narrow first-build amendment is independently source-reviewed (candidate995bc8b2, review8ca359c7). It removes the requirement to predict a Cargo output hash or native compiler probe outcome before observing the first real build. Each build-script unit instead needs a source-reviewed finite list of complete, correlated effects; linked paths must refer to that same package's actual canonical OUT_DIR. Unknown or duplicate script records are refused.

    This preserves the existing boundary: capture the full actual graph, script JSON and both executable hashes, then stop for independent exact authentication and separate root admission before formatting or projection. Later exact graph checks, all17 expectations, R3 ordering and final proof/CI requirements are unchanged. No discovery build, extra all-host executable build or factory implementation was introduced.

    This is source preparation only. The templates remain disabled: concrete registry/native effects, actual tool/environment/custody/resource controls and capacity are still unresolved. No checkout adoption, build, probe, projection, test, cleanup or identity writer ran. Superseded candidates and all #316 evidence remain preserved; #405/#467 stay with their existing owners.

  6. afflom commented on Nov 12, 2026

    @afflom
    MemberAuthor

    Factory preparation update (2026-11-12 09:41 UTC): the remaining seven registry build scripts and local compiler build script now have independently reviewed source descriptions. The registry revision85ffa9a8/review9ea01335 clears finite direct-effect alternatives for proc-macro2, libc, serde_core, serde, rustix and camino under their stated controls. Libm's environment-derived values remain explicitly unresolved. Local compiler review9a95e645 confirms three output files and the existing905-input boundary, reusing its retained inventory without another probe or hash sweep.

    Review removed two unnecessary proposed changes. Existing V5 already binds Serde's private.rs to the same package's actual output directory; no new driver mechanism or predicted Cargo hash is needed. A separate gate solely for embedded header/order formatting is also unnecessary under the existing exact producer/source/input controls and mandatory actual-output review. Those acceptance obligations remain intact.

    Concrete remaining work is now controls and output custody: finite libm feature/order/profile/target values; proc-macro2/rustix compiler flags, wrappers and child-output residuals; native/tool provenance, including libc's unconditional emcc version attempt and conditional freebsd-version invocation; actual candidate roots/OUT_DIR/graph/BOTH authentication; and capacity. All descriptions remain disabled source data, not admitted execution or delivered factory behavior.

    Completed source custody remains reusable:88 cached archives equal committed v0.5 lock checksums and4,048 files equal archive bytes; only exact Cargo unpack markers are excluded. Prior six-script, quote and BLAKE3 work was reused. No build, test, probe, writer, cleanup or identity wave ran.

    #476's bootstrap transfer stays with its owner: old-head55/55 is scoped evidence, while current8ab6 retains the environment-guard finding and lacks full qualification. The factory will consume accepted work rather than duplicate that producer. #405/#467 ownership and all #316 evidence remain preserved. Build capacity and these controls still block execution; #468 and #128 remain open.

  7. afflom commented on Nov 12, 2026

    @afflom
    MemberAuthor

    Narrow source-only progress: libm's proposed controls now identify exactly two feature-vector orders (arch/default or default/arch) and OPT_LEVEL=0 for the existing development-profile command, conditional on the stated exact environment and absence of overrides. Candidate1a55b95b / independent review910c1ccc retain those qualifications; no extra driver mechanism or discovery build was added.

    The target-feature value and complete effect alternatives remain null/STOP. Historical fxsr,sse,sse2 is not evidence of the future effective compiler/target/configuration. Environment enforcement, target/tool provenance, remaining child/native controls, actual graph/output authentication and capacity still block execution. Existing V5 and all completed source custody remain reusable. No Cargo, rustc probe, build script, test, writer or cleanup ran.

    #483 has since landed its owner batch with successful full CI, but the transferred-export environment guard and three other source findings remain open in #489. Factory integration will preserve that repair boundary and the existing #405 owner lane. Separately, the two ready documentation proposals now each have55 materialized CI jobs; coordination6824891015 asks their owner to consolidate corrections before the next qualification cycle rather than duplicate another full gate. No current run was canceled or acceptance requirement waived.

  8. afflom commented on Nov 13, 2026

    @afflom
    MemberAuthor

    A concrete remaining child-control omission now has an independently reviewed source-only amendment. The existing V5 pre-build environment comparison omitted RUSTC, RUSTC_WRAPPER, RUSTC_WORKSPACE_WRAPPER, RUSTC_BOOTSTRAP and RUSTC_STAGE. The one-line candidate adds those five keys to the same exact-map guard, so an unreviewed parent override fails before Cargo; it adds no new mechanism or discovery build.

    Corrected candidate ff07ef47 / patch29659661 / independent reviewc29d7c4c are CLEAR for this source delta only. An initial truncated driver-hash binding was corrected and its superseded artifacts retained. The actual V5 driver9f6435a2 is unchanged. Parent environment checks are explicitly distinct from Cargo's generated script environment and effective compiler configuration; they do not authenticate those inputs by themselves.

    The proc-macro2/rustix child invocation descriptions reuse the completed registry source custody and finite effect reviews. Actual tool/config provenance, child generated-output custody, remaining libm/native controls and capacity remain unresolved STOPs. Capture and independent review of the actual graph/script outputs/BOTH before tool use remain mandatory. No build, compiler query, probe, test, writer, cleanup or source adoption ran, and #508 cross-run proof reuse is not admitted. #405/#316 ownership and evidence remain intact.

  9. afflom commented on Nov 15, 2026

    @afflom
    MemberAuthor

    Upstream status correction: PrismPM#586 merged into release/v0.4 on2026-11-12 as8ccb0f96d30928c4083f1970094149ee5af3b3c4. I corrected this issue's stale “unmerged” sentence. The actual tree589aa279 differs from proposal e8eb0ad1's tree193fdac8, so source adaptation must bind the inspected actual; this does not establish inclusion in the pinned v0.4.0 release or qualified LexLean reuse.

    The actual source supplies a useful division: Factory.Engine models obligations/identity material, Factory.Workflow owns the generated skeleton, and host actions fill job fragments and admit receipts. Its ten-obligation graph is PrismPM-specific. Receipt trust still depends on branch writers, and the inspected expiry helper does not itself reject absent/malformed expiry; LexLean must retain its stricter unavailable/unbound-evidence rejection rather than copy that policy.

    The bounded next source candidate is the correspondence from the already-reviewed V5 proposal/control receipts to one LexLean-owned obligation. Reuse those inventories by reference, leave unresolved provenance NULL/STOP, and keep V5 as the sole potential executor. This proposal does not enable routing, generate success/reuse receipts, install a scheduler, replace the driver or admit a build. Real Prism-controlled developer/CI/release routing, qualification and retirement of duplicate orchestration remain required for closure.

    Existing #467/#405 owners and all #316 frozen evidence remain unchanged. No producer, compiler query, test, build, cleanup or identity wave ran. #468 and #128 remain open.

  10. afflom commented on Nov 15, 2026

    @afflom
    MemberAuthor

    Completed the bounded source candidate described in6890210052 and an independent source review. The pure offline adapter binds14 existing V5/control references by exact supplied bytes and maps13 unresolved template fields to explicit null request slots. Closed schemas reject duplicate/missing/extra fields and changed bindings. The unchanged V5 driver remains the sole potential executor; no inventories or driver controls were duplicated.

    Candidate receipt edb00a5366d0c674231e7ddcc1f97dc4ce31f514be61214a8291af672af13e47; independent source review57cb15db1d1c6e3c4d89a94ad7af4e526d2b59d1f69a6edb79afc4eb7d29c5ad confirms the three candidate files,14 reference hashes and13 null-field mappings. Source returns only blocked, null identity/evidence and execution=false. It has no filesystem/process/driver/planner invocation.

    This is data correspondence preparation, not a verified Prism model, real routing, admitted runtime evidence, reuse, a scheduler, or delivered factory integration. No import, execution or test was performed. The fixture and focused refusal matrix are retained for subsequent bounded validation; actual tool/config/graph/script/custody/resource fields remain unresolved. Current full-acceptance, same-run honesty/reproducibility, actual-tree and ownership requirements remain intact. #468/#128 remain open; all #316/STOP evidence is preserved.

  11. afflom commented on Nov 15, 2026

    @afflom
    MemberAuthor

    Completed one bounded validation of the independently reviewed pure request adapter. The pinned fixture returned blocked, and all twelve malformed/unbound input cases produced their exact expected refusal. The run used14 pinned reference byte inputs, returned0 with empty stderr, and preserved the wrapper's28-field #316 guard. Independent terminal review verified the cases, pins and output hashes without rerunning anything.

    Terminal receipt SHA256 f77e3ba418e6407f1d48199f4f8897171ce9d05ecd722287cc3640fc70afa288; independent terminal review0e3101574f3f7eba9d8a99a9419e1ee93aedb48367ec5a73690c4bb59ad4aa63. The check ran once under the whole-phase validation lock using the fixed interpreter, durable session, DEVNULL and external logs. The prelaunch386351104 free bytes was only a point observation, not a capacity guarantee.

    This validates only the pure adapter's blocked/refusal behavior. V5 was read as pinned bytes and never invoked; no Cargo, writer, build, native provisioning, factory routing, identity/reuse receipt or model proof was admitted. The existing disabled driver remains the sole potential executor. Actual tool/config/graph/child/native/custody/capacity bindings remain the next unresolved integration work. Reuse this completed evidence; do not repeat the check or start a producer from this result. #468/#128 remain open; #316 and all STOP evidence retained.

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

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions