Skip to content

test(bootstrap): avoid discarded BS-15 full-library runs - #459

Merged
afflom merged 4 commits into
release/v0.4from
dot/457-bs15-render-materialization
Nov 14, 2026
Merged

afflom merged 4 commits into
release/v0.4from
dot/457-bs15-render-materialization

Conversation

@afflom

@afflom afflom commented Nov 10, 2026 •

Copy link
Copy Markdown
Member

Closes #457

Change

Assemble BS-15's existing ordered evaluator/native render vectors directly. The previous code materialized a full-library Run for every render, then excluded large libraries from the evaluator and discarded their full native inputs while rebuilding the existing 4,096-byte-prefix runs. The discarded full runs fed no differential.

For small libraries, keep the same render_run and the same supplied package or recipe outcome in both vectors. For large libraries, construct only the same bounded native input and recipe outcome. Ordering remains accepted manifests, shaped inputs, then seeded render mutants. On the base tree's asserted counts, 735 native render cases and 241 evaluator cases are retained; 494 unused full-library encodings disappear. That is a source-level work-count deduction, not a measured speedup.

Preserved

All corpus generation/seeds/counts, negative fixtures, complete accepted-manifest package::package/recipe assertions, lowered roots, 21 planted defects, native differential, certificate limits and exported-package byte/program comparisons are unchanged. Only conformance source and verification documentation change. No SPEC/model/compiler-semantics input, generated product, host gate, timeout, CI orchestration, cache or security/protection change.

#442 continues to own count-report/writer tooling. Coordination is posted there; retain its reporting/fast return when reconciling the nearby hunk. #272 remains open for the complete certified packager.

Validation

  • Independent source review found no blocker: old/new membership, ordering and expected-outcome routing are equivalent; removed calls only encode values or compute deterministic recipe bytes and do not perform unique verification.
  • Added library regression render_corpora_preserve_inputs_and_outcomes, retaining the previous assembler as reference. It covers both profiles, distinct supplied package outcomes, recipe outcomes and libraries of 0, 4,096, 16,383, 16,384, 16,385 and 32,768 bytes.
  • Local whitespace checks passed. Independent Python assembly model passed 27 empty/mixed collection cases and detected changed boundary, prefix and supplied-outcome routing. A separate reviewer also checked 100 randomized mixed cases. This is logical call-routing evidence, not execution of the Rust renderer.
  • Remote source readback matches the reviewed artifact byte-for-byte. Base: cf76090d473540f4008ec373d4c5ed96ad13cc01.

The authoring environment has no Rust toolchain. Hosted draft CI 42740747766 and reproducibility 42740747949 passed on head 5088c0f47d8a15abcb6a2567c857cfa79fd01b39. The rest job ran the actual new Rust regression successfully (44 library tests passed; none failed, ignored or filtered), and workspace Clippy passed with warnings denied.

After those runs completed, this PR was marked ready once to start full qualification 42747441605 on the same head and unchanged base. That run passed all 50 required jobs and every step on the earlier exact tree. Independent artifact review verified supported hosts, Intel custody and all 478 cases, Atlas, and actual BS-15 execution. Linux BS-15 took 6,090.03 seconds versus 10,304.13 seconds in the prior baseline, an observed 40.9% shorter run with matching verifier event counts, not a controlled causal benchmark.

At the final merge check, concurrent #395 had advanced release/v0.4 to c66f83d05bc7d488f6328dba6d1521c2e4520af4 and introduced a conflict. The old result does not qualify that new combination. Source reconciliation and contributor landing-order coordination are in progress; no merge or replacement full gate has been started. Target milestone is v0.4.1, followed by accepted forward integration into #405 before the v0.5 candidate freezes.

Assemble the same ordered evaluator and native corpora directly. Keep the
existing full accepted-package recipe assertions, bounded native inputs,
all plants, counts and exported-package checks. Add an old/new assembly
regression without changing compiler semantics or gate orchestration.

Refs #457

afflom commented Nov 10, 2026

Copy link
Copy Markdown
Member Author

Release-line clarification: this is post-release v0.4 maintenance, not evidence for the original v0.4 release. Keep release/v0.4 as the current qualification target. A final landing must also advance through the #405 forward-integration lane so upcoming releases receive the same test-only improvement. The PR remains draft; no full gate has been requested.

afflom commented Nov 10, 2026

Copy link
Copy Markdown
Member Author

Assigned this fix and #457 to milestone 3, LexLean v0.4.1, matching the existing next-patch plan on release/v0.4 after published v0.4.0. This is upcoming patch work, not preparation for the already-released v0.4.0. After complete acceptance and landing, it must also be integrated through #405 into release/v0.5. The broader v0.4.2/#272 product scope remains unchanged.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent source review found no blocking defect in the render-corpus transformation. Accepted → shaped → mutant order is preserved. Libraries at or below 16,384 bytes retain identical arguments and outcomes in both consumers; larger inputs retain the same native-only 4,096-byte prefix and recipe outcome. accepted() still compares the complete library with package::package before assembly. Seeds, counts, negative fixtures, all 21 plants, evaluator/native differentials, certificate budgets, eligibility and export comparisons remain unchanged.

The regression reference reproduces the former assembler, with distinct package outcomes to detect accidental oracle substitution. Both profiles and six boundary sizes are covered. Read-only assembly-model checks cover empty/mixed inputs, randomized inputs and rejected boundary/prefix/oracle-routing mutations; these are not Rust execution evidence. The follow-up commit only applies the exact rustfmt assertion wrapping. Final source SHA-256 5abc47a228f6c9a47dec653e4cb99b06bef2bcdad254d466630ffde86ef824f1 matches remote blob 745fcae6969878566352009ed9d0843c776cf0c8.

This removes 494 full-run encodings that neither consumer used, but does not establish elapsed-time savings. Source identity inputs and products are unchanged. Hosted final-head honesty has passed; rest and reproducibility are still in progress. Actual library regression, BS-15, supported-host and complete final acceptance remain required. This COMMENT is source review, not readiness or merge acceptance.

afflom commented Nov 10, 2026

Copy link
Copy Markdown
Member Author

Draft qualification is green on exact head 5088c0f47d8a15abcb6a2567c857cfa79fd01b39, base cf76090d473540f4008ec373d4c5ed96ad13cc01:

  • CI run 42740747766: honesty and draft rest passed, including pinned formatting, model/spec/meta checks, workspace/all-target/all-feature Clippy with warnings denied, tests/features and inventory.
  • Rest job 129582053241 explicitly executed cases::bootstrap::rust_package::tests::render_corpora_preserve_inputs_and_outcomes ... ok; the library result was 44 passed, 0 failed/ignored/filtered, in 1.80s.
  • Reproducibility run 42740747949 passed for the same exact head/base.

The initial draft run failed only on the new unit assertion's rustfmt wrapping. The follow-up applied exactly that whitespace change; independent final source/readback review retained the original equivalence finding.

This is development qualification, not the complete release gate. conformance_bs_15 did not execute in draft rest, and the full BS-15/planted/native/export checks plus complete qualification remain required. No measured CI speedup is claimed. The PR remains draft for coordinated v0.4.1 qualification; the accepted landing must also be carried forward before the v0.5 freeze.

afflom commented Nov 10, 2026

Copy link
Copy Markdown
Member Author

Full qualification on head 5088c0f and base cf76090 completed successfully: https://github.com/UOR-Foundation/LexLean/actions/runs/42747441605. All 50 expected jobs and every reported step passed; exact tree and supported-host/Intel/Atlas evidence were independently checked. Linux BS-15 took 6,090.03 seconds versus 10,304.13 seconds in the prior baseline, an observed 40.9% shorter run with matching verifier event counts, not a controlled causal benchmark.

The final merge check caught the concurrent #395 landing at c66f83d, which adds BS-21 and changes the corpus/gate graph. This PR is now conflicted, so the completed gate is evidence for the earlier tree, not permission to merge the new combination. Preparing a minimal reconciliation preserving both histories and the new accepted products; no protections bypassed, no products regenerated, and no redundant gate started. We will coordinate the next candidate with the contributor queue before requalification.

Retain accepted release/v0.4 source, products, corpus counts and gate plans
from c66f83d. Preserve the reviewed BS-15 assembler and regression. Keep the
accepted BS-21 verification history and distinguish historical/current
work-count deductions. Prior-tree qualification is historical evidence;
this combined candidate requires its own complete current-graph gate.

Refs #457

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Independent bounded reconciliation review of exact a342e22a against actual release c66f83d0: no blocking source defect found. The ordered parents are prior candidate 5088c0f4 then actual c66f83d0. Only the BS-15 conformance file and VERIFICATION.md differ from the accepted release; BS-21 source/products, corpus additions and gate plans are preserved.

The direct assembler and boundary/oracle-routing regression are byte-identical to the previously reviewed candidate. Current accepted counts retain 737 native renders and 241 evaluator renders, avoiding 496 discarded full-library encodings. The verification record distinguishes these from the earlier 735/241/494 deduction. All seeds, plants, full-library package assertions, native/evaluator membership and export checks remain intact. This is source-level work reduction, not a new timing measurement.

GitHub's current synthetic merge 3ea03613 has the same 975f536e tree as the reviewed head. Exact-head repro42764267541 passed; full current-graph CI42764267561 is still running, so this COMMENT is source review only. Earlier 5088/cf760 results do not qualify this combination. Complete final-head acceptance, owner final review and actual merged-tree equality remain required before #457 closure; #272 stays open and actual landing must be forwarded through #405.

The PR body still describes reconciliation as pending and retains the old work-count context; please update it for the final candidate when recording qualification. No local producer, test, log retrieval or replacement gate was started by this review.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Final-head CI acceptance review for exact a342e22a / actual base c66f83d0: CLEAR for this candidate. CI42764267561, attempt1, completed all 51 dynamically materialized jobs and all 564 reported steps successfully. This includes same-run honesty, the current BS-21 structural-grammar job, BS-15 package, every supported host, Intel construction/partitions/aggregate and the Atlas join. Exact-head repro42764267541 also passed. No earlier/canceled run supplies any part of this acceptance.

Independent review rechecked the whole #457 body against exact-head source review5837900888. Ordered vector equivalence, boundary/prefix behavior, regression coverage and preserved seeds/counts/21 plants/native/export guarantees remain satisfied; the current source avoids 496 discarded encodings while retaining 737 native and 241 evaluator renders. No new timing claim or benchmark threshold is introduced. Existing #395 count-description findings remain separately tracked in #412.

Head, CI-recorded tree, repro-recorded tree and current synthetic merge 3ea03613 all bind 975f536e8974c64d2851b19815f234d8eec3a845. This review uses live job/step metadata and retained source evidence; it does not claim a new raw-log or artifact-content audit. No producer, test, log download or gate was repeated.

The existing owner retains final qualification and landing. Please update the stale PR description to the current reconciliation/qualification when recording that decision. After an actual merge, verify its tree equals 975f536e before explicitly closing #457 with this merged PR link. #272 remains open; carry the actual accepted landing through #405 before the v0.5 freeze. The synthetic merge is not evidence that delivery has happened.

afflom commented Nov 10, 2026

Copy link
Copy Markdown
Member Author

Landing-order update: #451 was externally merged at16:59:07 UTC as700196ceadd9695abd47bee29b94a6b8bec6f915 (actual c66f+441338 parents, reviewed/tested c0148ddd tree). Your a342 head remains unchanged and unmerged. Its completed51-job qualification and final review remain valid evidence for the c66f/a342 tree975f536e; they do not establish acceptance of the newly combined tree.

Existing ownership stays with you. Please reconcile against actual700196 history and review the resulting source/product impact before deciding the necessary qualification; no automatic rerun, duplicate patch, producer or root takeover has been started. The completed gates/logs remain preserved. #457 stays open pending accepted actual landing and full-body delivery; #272 remains open and accepted history must reach #405.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded current-release integration review for a342e22a7fddf158d9e3ed948741379d733b72b9 against actual release/v0.4 at bc71b33f9a4ce7f5582c337b5752a7ba9eb49706:

  • #451 changed only three constants in crates/conformance/src/cases/bootstrap/rust_package.rs: manifests 319 → 324, accepted packages 300 → 305, and render mutants 300 → 305. Those hunks are separate from this PR's assembler, call and regression. #466 leaves that entire Rust file identical to #451 (blob 1541366c2dc2a6d1c23ae53a9bb1792d9843b1f0).
  • Retaining the accepted corpus gives 747 native renders, 241 evaluator renders, and 506 discarded full-library encodings avoided. These are source deductions, not new timing results. The existing 737/496 and 735/494 records remain valid for their historical bases.
  • The verification-history additions are also in separate locations. Preserve the accepted #451/#466 history and current products when reconciling; the two-path optimization itself indicates no need for another identity or product writer.

This is an independently reviewed source-impact assessment, not a composed-tree acceptance or a merge-conflict guarantee. Existing ownership remains unchanged. The next candidate should retain actual bc71 ancestry and current constants, state the current counts, and receive exact-resulting-tree review followed by the required complete final-head qualification. Prior completed gates and measurements need no baseline repeat. #457/#272 remain open pending their respective full delivery scopes; any accepted landing must reach #405.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded integration review credits afe6c67's ordered merge of reviewed a342 and accepted actual fd7. Net source remains the existing BS15 assembler/regression plus verification history; no new implementation defect was found. Current corpus constants imply797 native renders,241 evaluator renders and556 unused encodings avoided. These are source deductions, not new timing measurements; historical counts and timings retain their original bases.

Please update the historical PR description with this composition and the current qualification. Existing CI43493536296 is queued and exact repro passed; prior51/51 acceptance is scoped to a342's earlier tree. Actual release has since advanced through #517/be7 and #462/6a02 (automation and counts-tool changes). Review that narrow incoming history and final resulting-tree applicability rather than starting an identity writer: none is indicated by this patch. Finish the existing required final-head gate and review before landing, then verify actual tree and explicitly close #457; #272 stays open. No old checks, logs, producers or measurements were repeated.

@afflom
afflom merged commit f8854bf into release/v0.4 Nov 14, 2026
71 checks passed
@afflom
afflom deleted the dot/457-bs15-render-materialization branch November 14, 2026 00:12
humuhumu33 added a commit that referenced this pull request Nov 14, 2026
Brings the combined batch current with release after #459 and #498 landed,
before its regeneration runs.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
afflom pushed a commit that referenced this pull request Nov 14, 2026
…urity-patch

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Postlanding acceptance of actual merge f8854bf, tree90732a41fa749a83859577c2ea3fb52df343e6b6: exact push CI43532545468 attempt1 completed all70 jobs and772 steps successfully, including same-run honesty, BS-15/package, native hosts and Atlas/Intel/Windows joins. Exact actual reproducibility43532545412 also succeeded. Actual commit/tree and checkout binding were verified; the prior afe/b539 proposal gate was not substituted for this different merged tree.

Independent full-body #457 acceptance is clear. Retained source reviews establish ordered old/new vectors, <=16384 boundary,4096prefix and both-profile regression while preserving corpus, seeds, counts, assertions, negative fixtures, two roots,21 plants, native differential, certification limits and export-byte comparisons. No repeated test/build/producer was needed. Receipt: issue457-final-acceptance-20261114/receipt.json.

This completes narrow #457 after landing; it is not retrospective premerge approval or a measured current speedup claim. The PR's original reconciliation/timing narrative is historical. Broad #272 stays open, and actual accepted forward integration remains separately owed in #405.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant