Repository navigation
BS-72: the reasoning and text declarations and whole semantic modules as source (SemanticReasoning) - #497
BS-72: the reasoning and text declarations and whole semantic modules as source (SemanticReasoning)#497humuhumu33 wants to merge 45 commits into
Conversation
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
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
…al declarations) One closed declaration that binds a foreign function to a stated contract: an opaque constant whose hidden value is a LexLean witness, and one generated contract axiom that every dependent theorem carries in its exact observed set and verify reports with its evidence. Options, a Lean 4.32.1 prototype with its output, an assessment of an Aeneas bridge, and the decisions left to the owners. Proposal only: nothing is implemented. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A shard writes toolchain.json (lexlean/verify-toolchain/1) beside its output: the SHA-256 of the lean, lake and leanchecker its preflight hashed, of the running lexlean, and the address of the rustc that compiled it (its executable and rustc_driver library, hashed by the build script). A part that reads another refuses it (LLI9001, naming the part, the tool and both addresses) unless all five equal its own. A process record names only lean or leanchecker, so a part that ran another lake, lexlean or rustc was accepted before. VR-23 plants another lake address in a shard's output and, on unix, runs a join whose own lake differs from the pinned one by one byte; both are refused. Published directories, attestations and expected/verify do not change, and neither does the compiler-semantics ID. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… as source The logic, inference rule, verifier, reasoner, tokenizer, context window, text generator and text classifier declarations (SPEC.md §17.12) with the eleven types they hold, and SemanticModule of declarations of every kind, as language-1.2 source in bootstrap/, eighth slice of the typed IR (#269): reasoningDeclarationText and semanticModuleText of SemanticReasoning, held to the canonical JSON of their serde form for every committed semantic module. Generated with the case encoders and examples from one description of the oracle types, as BS-71 is; 21 planted defects. Register row (501), SPEC §31 row and §17.23 prose, scenario, bdd part and lane, README count and range, CHANGELOG, VERIFICATION record, roots, the register-count mirrors; BS-71 case helpers made pub(super). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The custody record becomes lexlean/ci-bootstrap-export/2. Beside what `rustc -vV` and `lean --version` report, it names the SHA-256 of the pinned toolchain's lake, lean and leanchecker and the address of the rustc that compiled the exporting binary (SPEC.md §22.11). A leg whose own executables differ refuses the export. A version banner is the same for a rebuilt or patched binary; the bytes are not. RP-14 plants another leanchecker and another rustc. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lders Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The stack-bound step of the regeneration maps a package to its root by name, so a root is the camel case of its package and unique in the project: semanticTermText, semanticProofText, semanticDeclarationText, semanticModuleText, and wholeSemanticModuleText. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-addresses # Conflicts: # CHANGELOG.md
…ning Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ning Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Separate proof-axiom dependencies from executable external uses (call inventory), complete the report-to-implementation binding, narrow the opacity statement. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
…, record D1 One reviewed documentation carrier, as requested on #492. D1 (SPEC.md 9.4 reuse of verification results) is decided: not adopted now. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
…ning Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…(BS-50..52, #269) One batch branch carrying three independent slices of the lexicon closure, as BS-25..34 landed in #480: - BS-50 LexiconSchema: the category/frame/channel/associativity/feature tables and surface_safety/is_lean_name/toml_comment_at/entry_path, over the atoms of BS-03. Six roots. - BS-51 LseScopes: Lse::check_scopes, referenced_consts, and outer_explicit_binders/result, over the expressions of BS-25. Three roots. - BS-52 LseTypes: the conservative checker, check_signature and check_value, structural over a host-given nesting bound. Two roots. Register rows (model/ids.toml, SPEC §31, VERIFICATION.md, features scenarios) are inserted as one whole block per slice, in register-ID order, after BS-34. Counts and totals that the regeneration itself writes (README.md, CHANGELOG.md, the §31 total line, repository.rs's and spec_links.rs's row counts) are left at release's current values for the regeneration to fix. .github/workflows/regenerate.yml is release's version overwritten with the fixed one from ci/regenerate-relocks (the deliver job and a relock step that survives combine), so this branch's own regeneration lands itself onto the PR opened for this batch branch. #522, #523 and #524 stay open as the per-slice record; they are not what lands. Branched from release/v0.4 at 6a02d2c. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
SPEC.md's "Total required capability IDs", xtask/src/spec_links.rs's RP-07 check, and repository.rs's RP-07 assertion still said 496: none of them is written by a recipe (cargo xtask counts --write carries the counts a bootstrap case reports, not the §31 row count), so the finish job would not have fixed them and RP-07 would fail after the whole regeneration. All three now read 499, matching model/ids.toml's 499 rows and SPEC §31's 499 rows. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Carries all eight completed grammar/surface slices (GrammarProof, SemanticPrint, SemanticRead, CoreJson, CorePrint, CoreRead, GrammarMath, PropositionTerms) as one batch, as BS-25..34 landed in #480. Hand-authored register content (SPEC.md Sec31 rows + Sec17.23 prose, model/ids.toml entries, features/suites/bootstrap.feature scenarios, VERIFICATION.md, bootstrap/README.md, bootstrap.rs dispatch wiring, bootstrap/lexlean.toml entrypoints, the two xtask/repository.rs register-count constants, and .github/bdd-parts.json, vv-rows.json, intel-host-lanes.json CI-plan entries) is unioned in register-ID order onto release's current content. Regeneration-written files (CONFORMANCE.md, ERRORS.md, roots.json, lexlean.lock) and README.md/CHANGELOG.md are left untouched -- release's side -- for the one regeneration run to fill in. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ialization test(bootstrap): avoid discarded BS-15 full-library runs
Merges the grammar batch (cd571f1: BS-35..40, 42, 43) and the lexicon batch (ebe8911: BS-50..52), both authored directly on release 6a02d2c, into one combined batch so neither lands alone and makes the other stale. Conflicts at the shared BS-34 insertion anchor are resolved as "keep both, grammar then lexicon" for the register-order files: SPEC.md §31 rows and §17.23 prose, model/ids.toml, features/suites/bootstrap.feature, bootstrap/README.md, crates/conformance/tests/conformance.rs, and crates/conformance/src/cases/bootstrap.rs. Three of those unions dropped a syntactic separator at the merge seam (ids.toml a missing [[id]] header, conformance.rs a missing closing brace and #[test] attribute, SPEC.md an extra blank line) -- fixed by hand and reverified: 507 unique ids, 507 §31 rows, 42 feature scenarios and 42 conformance fns, no stray conflict markers. Hand-merged rather than unioned: bootstrap/lexlean.toml entrypoints (41+3=44), the register-count constant repeated in both xtask/src/spec_links.rs and crates/conformance/src/cases/repository.rs (bumped 504/499 -> 507 at every site), .github/bdd-parts.json and vv-rows.json (all eleven new parts and rows), and .github/intel-host-lanes.json (grammar keeps its chosen lanes; the three lexicon parts land in "lower", matching where #525 put them). bootstrap/lexlean.lock keeps release's side (the lexicon batch dropped it; the runner relocks regardless). .github/workflows/regenerate.yml stays byte-identical to ci/regenerate-relocks. Supersedes #525 and #527, which stay open as the per-batch record. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
verify: every part names its toolchain by address (UOR-native, piece 1)
…ntracts docs: proposals for external contracts and for reused verification results
…nding PR #497 becomes the landing PR for BS-65..72 (the typed-IR stack), stacked on #532 (the combined lexicon and grammar batch, BS-35..40, 42, 43, 50..52). BS-73 (SemanticDecode) comes off the critical path: its certificate needs more than 900 seconds of lean time on the runner and twice killed the runner outright, so the other eight slices land without it, on their own branch and PR #504. Register: 515 ids in order (BS-35..52 before BS-65..72), bootstrap/lexlean.toml entrypoints the union of both (44 batch + 8 ours), child_timeout_ms and max_file_bytes as release has them (no BS-73 content anywhere), and regenerate.yml byte-identical to ci/regenerate-relocks (PR #521s fixed relock step). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
Bounded integration review at9d276 follows the body's selection of #497 as the BS65–72 landing carrier over #532, with BS73 remaining separate in #504. The complete54-path delta from a5dc adds the combined grammar/lexicon and actual69de history, preserves prior IR case blobs, and introduces no generated-product changes. The shared-lock publication repair is credited: all writers refresh working locks, while production-export alone publishes the three shared lock deltas. This reuses the #521/#504 review; no extra workflow test or producer is requested.
P1 from #532 review5901581394 is inherited here: tests/production-export/bootstrap/roots.json (4ea063a9) retains95 lexicon/release/IR roots but omits the same15 grammar roots. The52 entrypoints do not supply those production-export request rows. Correct the selected combined input and disposition applicable official outputs before treating regeneration as complete coverage. BS35/36/37 universal-versus-sampled claims and the unchanged BS67 corpus/fallback and BS69 evaluator-coverage findings also retain their prior scope; integration does not resolve them.
Align the PR base/description with the intended landing sequence: live base still names #496's semantic-models branch while the body says stacked on #532 and selected for landing. Preserve actual forward history and avoid a competing carrier. The assertion that BS73 needs >900s and caused runner deaths remains owner attribution; retained #504 evidence establishes interruption, not that causal diagnosis.
Current regeneration43558647460/CI43558650466 are pending and repro43558650447 failed, with no cause inferred here. Existing owner retains implementation; qualify the corrected final tree and verify actual landing equality. Receipt pr497-9d276-integration-delta/receipt.json. No case, log, producer or gate was repeated.
All eight new bootstrap-{ir-term,ir-proof,ir-document,semantic-types,
semantic-terms,semantic-decls,semantic-models,semantic-reasoning} parts
had landed in print (already the heaviest lane at 20), leaving lower 12,
print 28, validate 11, moves 6. Moves five of the eight lighter/IR-layer
parts (ir-term, ir-proof, ir-document, semantic-terms, semantic-decls) to
moves, two (semantic-reasoning, semantic-types) to validate, and one
(semantic-models) to lower, spreading what is believed the three heaviest
cases (semantic-reasoning, semantic-models, semantic-terms) into three
different lanes rather than stacking them. New totals: lower 13, print 20,
validate 13, moves 11. Every part release or #532 placed is left alone.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Correcting my own earlier instruction: these two files are not regeneration output -- nothing writes them, every slice PR edits both by hand, and the gate's RP-07/RP-11 checks read them. README.md's count sentence and the bootstrap capability row's BS-01..BS-34 claim, and CHANGELOG.md's Unreleased bullets, were all left on release's side in PR #532 and would fail the gate once its regeneration delivers. README.md: the count sentence becomes "All 507 registered conformance IDs ..."; the bootstrap row's range becomes BS-01..BS-52; and the row's prose gains what each of the eleven slice PRs (505, 509-515, 522-524) added to it, in register order, taken verbatim from each PR's own patch against its merge base. CHANGELOG.md: every bullet each of the eleven PRs added under ## Unreleased, in register order, verbatim (including each independent slice's own max_child_output_bytes bullet, since each PR added it as its own change at authoring time). Built as a separate branch on top of batch/bs35-52's current head rather than on it, since #532's regeneration (an hour in) must not be disturbed and neither file is an input to anything it generates. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
README.md and CHANGELOG.md are not regeneration-written, so they needed the BS-35..52 prose and changelog bullets that #532 (batch/bs35-52) was missing. Keeps 515 registered conformance IDs and the bootstrap row's range at BS-01..BS-72, with both sides' clauses in the row's prose (the fix branch's eleven BS-35..52 clauses, then this stack's own, in register order) and both sides' changelog bullets under Unreleased. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Attempt 1 and 2 of r16's finish job both lost the runner outright in counts and stack bounds, neither hitting the steps own 360-minute timeout: The hosted runner lost communication with the server. BS-69, 70 and 72 (semantic_terms, semantic_decls, semantic_reasoning) were still running together, at the default test-threads (4 on this runner), when each attempt died; a single-file verify of a similarly large semantic-module case measured over 6 GiB RSS alone, so four of them at once on a 16 GiB runner swaps the agent to death rather than finishing or timing out cleanly. Splits the already-computed case list into a known-heavy set (BS-69, 70, 71, 72) run one at a time (--test-threads=1) and the rest run at --test-threads=2, instead of one cargo test invocation at the default thread count. For this landings eight cases (BS-52, 66-72), expected total: roughly 60 to 90 minutes for the light four at threads=2 (they ran four-way in about 2 hours before; halving concurrency should cut that substantially, though not precisely predictably), plus the heavy four serialized, each guessed at 15 to 40 minutes from BS-71s one completed run (14.4 min under partial, not zero, contention) -- a rough total of 140 to 250 minutes, comfortably under the steps 360-minute cap even if that guess is off by a wide margin. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
Bounded one-path workflow delta: the existing computed case list is preserved, with BS69–72 running at one test thread and remaining cases at two; no product adoption or new case/source qualification is established here.
Please qualify the new regenerate.yml:314–317 resource comment or link the already-retained evidence. It says EACH of BS69–72 measured over6GiB RSS and that four concurrent cases 'swapped the agent to death twice'; the commit instead cites one similarly large semantic-module measurement and runners losing communication. Those observations do not establish four per-case peaks or the cause of agent loss. State the concurrency choice as a planning assumption unless existing receipts prove the stronger claims. Reuse those receipts; no new benchmark, probe, producer or failed-run repeat is requested.
Separate producer972c7bb is not this exact9f064 head, so do not treat its progress as current-head acceptance or adopt its outputs without explicit input/output binding. Existing owner retains source corrections and final qualification. No runtime or stopped#431/oldTC03 review was performed.
|
Current exact9f064 regeneration43616108537 is terminal canceled: the three writers and combine succeeded, counts/stack-bounds job132552791121 was canceled, and deliver was skipped. This does not establish final output or full acceptance, and cancellation alone does not establish OOM/agent-loss cause. Please retain those completed upstream results and identify their input/product custody plus the counts-stage disposition before choosing a necessary continuation. Avoid restarting all three writers merely because the downstream stage stopped. This is not permission to aggregate canceled work into success or adopt outputs from another head. Existing coverage/resource-attribution findings and implementation ownership remain; no root retry, new producer or CI control was initiated. |
…532) Replaces .github/workflows/regenerate.yml with the version prepared for #532 (blob 704df4f): heavy cases (BS-36, 69, 70, 71, 72) run one at a time with explicit started/ended lines and elapsed minutes, the rest at --test-threads=2, and the whole step is bounded by a shared timeout against a 320-minute budget so a run that cannot finish everything fails cleanly (and keeps every row already measured in tests/stack-bounds.json) instead of being cancelled by the jobs own 360-minute limit and losing all of it. Also carries the once-a-minute free-memory trace from the same change. No other change; not a release merge. This copies ahead of #532 landing so #497s own stack-bound rows can start banking now, since measuring all of BS-65..72 is the longest pole and the two PRs rows stay valid through a later merge as long as the packages do not change. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
Bounded one-file workflow review at806f: case partition is retained, BS36 joins the individually serialized heavy cases, and pipefail plus overall_status keeps a failed/timed-out/exhausted measurement from becoming success. A failed finish can upload diagnostic partial evidence but still blocks deliver. No proof or whole-generation acceptance follows from that partial artifact.
Two limits need owner disposition before describing this as guaranteed timeout-safe salvage. The320-minute deadline starts inside measurement, after earlier finish-job work; the360-minute job limit includes that earlier work and return/upload. Derive a job-wide deadline with a return/upload reserve, or narrow the guarantee to best effort with the required earlier-step bound. Also, timeout has no --kill-after escalation: TERM-resistant processes or inherited pipeline descriptors can outlive the remaining budget. Add bounded escalation with its grace accounted for, or qualify the hard-bound claim. These are static conditional failure modes, not an assertion that this run encountered them.
Minute host-free snapshots do not measure per-case peak RSS or establish an OOM cause. Reuse completed rows only after binding their exact source/export roots, measurement implementation/options/toolchain/oracles and artifact provenance; unchanged package names alone are insufficient. Existing497/532 owner retains that disposition and the active producer43746749297. No duplicate writer, rerun, local probe or CI control requested. Receipt ae60d0a4b6ad0e94183457f1f641a1e09dc039ef71eca09ec4c975c8fb21260b. COMMENT only; current CI/repro failures and final-head acceptance remain separate.
afflom
left a comment
There was a problem hiding this comment.
Regeneration 43746749297 at 806f207 has failed downstream of successful writers and combine. The new counts log establishes exhaustion of the 19,200-second measurement budget: BS-52/66/67/68 passed in 12,625.49 seconds, BS-69 passed in 5,443.64 seconds, BS-70 began with 1,130 seconds remaining and ended without a success summary, and BS-71/72 were explicitly skipped for lack of time. This does not establish OOM or a proof failure.
Return/upload succeeded, preserving diagnostic artifact 14191847002 (digest d0e474189df02ba4243f2db1c391351c61e0bb9a8d0f18b9c43e862567dee2cd). Deliver was skipped. Artifact metadata and partial results do not establish completed delivery or acceptance.
Existing owner: preserve the successful upstream products and failed-stage evidence, identify the remaining BS-70/71/72 measurements, and specify the smallest bounded continuation with time reserved for saving results. Bind any reused rows to their exact source, package, toolchain, options and oracle inputs. Do not repeat all writers because this downstream stage exhausted its budget. The earlier timeout/salvage guarantee findings remain; no retry or execution is initiated by this review.
Receipt: /workspace/work/pr497-43746749297-terminal-20261116/receipt.json; SHA256 4ea9bb103f9e44e4e0916597d2f9e62155fb575cdaad999640117d7df0b75521. New job log 132946908728 retained once.
tests/stack-bounds.json only, taken from r18's (43746749297) regenerated tree and checked against this heads own copy first: exactly eleven rows added (bootstrap/check_signature, check_value, declaration_text, linked_text, module_text, project_semantic_id, proof_text, semantic_proof_text, semantic_term_text, type_spelling, type_text), zero removed, zero changed, same spec/headroom/granule/scaling, same sorted row order and pretty-printed format. The new blobs sha (1402fa8) matches the git-diff index line r18s own log recorded for this file, not just a JSON-level comparison. r18s measuring step hit its own 320-minute budget (by design, not the jobs 360-minute cancel) after BS-69 finished and BS-70 was cut off mid-run; BS-70, 71 and 72 still need measuring. Banking these eleven now means the next regeneration only has those three left. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
afflom
left a comment
There was a problem hiding this comment.
The one-file eee6add patch preserves eleven completed stack measurements: each added value matches the maximum recorded in the already-retained counts log. That evidence was reused without refetching or rerunning it.
The workflow is unchanged. Its step-local 320-minute budget, missing job-wide result-saving reserve and termination limitations remain; BS70 was unfinished and BS71/72 skipped. The rows do not turn the partial diagnostic artifact into completed delivery. Bind reusable measurements and writer/combine products to exact inputs and custody before the smallest necessary continuation, preserving time to save results. No blanket writer repeat is requested.
Current CI/repro failures are recorded as metadata only; no cause is inferred. Existing owner retains completion and base/body alignment.
Receipt: /workspace/work/pr532-497-owner-repair-deltas-20261116.json; SHA256 24812521642be3fba43aa5d31a129971d5e2a2e599b95a65fd3aff16f7c4c076.
afflom
left a comment
There was a problem hiding this comment.
Run 43834332532 now has a concrete BS-72 failure at unchanged eee6add. The new counts/stack log records BS-70 passing in 5535.29 s and BS-71 passing in 4412.29 s, then BS-72 panicking at semantic_reasoning.rs:1011 after 4958.08 s. At this exact source, line 1011 is the largest-certificate-part budget assertion. This is a certificate-size refusal; the available log does not establish the offending root or measured size because the workflow's grep filter drops both the BS-72 diagnostic and assertion message. It is not evidence of a timeout or OOM.
The failure salvage path worked: regenerated-eee6addb artifact 14239474203 was uploaded (103240242 bytes; SHA256 08182a3167beb42d224c76129f02f6e46d67faee6b323981a57b726243ff492e). This is metadata custody, not inspection of its ZIP contents. Delivery was skipped. Preserve the three successful producer parts, successful combine and completed BS-70/71 evidence.
Existing owner: use the retained artifact/products to identify the oversized part and make the smallest repair while keeping the specified budget. Preserve full failure diagnostics in future admitted work so a filtered log does not force expensive rediscovery. No blind rerun, producer repeat or budget increase is requested. No root execution was started; the draft remains unqualified.
BS-72is the eighth slice of the typed IR of #269, stacked on #496 (BS-71). It covers the logic, inference rule, verifier, reasoner, tokenizer, context window, text generator and text classifier declarations (§17.12), with the eleven types they hold. It also coversSemanticModulewith declarations of every kind. They are language-1.2 source inbootstrap/src/SemanticReasoning.lex.tex.reasoningDeclarationTextSemanticDeclarationvariantssemanticModuleTextSemanticModule(any declaration kind)With this slice, source writes every semantic module the linked IR embeds. Like BS-71, the module and the case's encoders and examples come from one description of the oracle's types.
The case is
crates/conformance/src/cases/bootstrap/semantic_reasoning.rs. Its runs:It also has 21 planted defects.
This PR is now the landing PR for the typed-IR stack,
BS-65..BS-72. It is stacked on #532 (the combined lexicon and grammar batch,BS-35..BS-40,42,43,50..52).BS-73(SemanticDecode) comes off the critical path: its production root's certificate needs more than 900 seconds ofleantime on the runner and has twice taken the runner down outright, so the other eight slices land without it, on their own branch and PR #504.State: draft, for the parent's batch.
lexlean checkandfmt --checkpass.case/runner branch.bootstrap/lexlean.tomlentrypoints are the union of both stacks;child_timeout_msandmax_file_bytesas release has them;regenerate.ymlbyte-identical toci/regenerate-relocks.🤖 Generated with Claude Code