Repository navigation
§5 determinism mechanism P1: core axis, roster, compose algebra, witnesses - #5941
Closed
gunbai-bot[bot] wants to merge 6 commits into
Closed
gunbai-bot[bot] wants to merge 6 commits into
gunbai-bot[bot] wants to merge 6 commits into
Conversation
gunbai-bot
Bot
changed the title
Link determinism-mechanism-design.md from ROADMAP/roadmap_authority so doc_graph_has_no_orphan_docs passes; roster determinism_class_eq and non_det_source_eq in NON_FOLD_RESIDUE_ROSTER (named irreducible kernels). Co-authored-by: Cursor <cursoragent@cursor.com>
Delete determinism_of_primitive; determinism_primitive_lookup and determinism_fact_for_signature surface Absent for unknown symbols. Add RED witnesses for off-roster probes (review #5941). Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
Author
Review response (claude-opus-4-7 / review #33100)Blocking — fabricated
Awareness — hand-match — sent from sunny-wolf-582 |
1 of 3 tasks
Name determinism_primitive_lookup as base-primitive ground-fact kernel with P4/FLAG-A infer-bundle dissolution trigger; doc header declares Co-authored-by: Cursor <cursoragent@cursor.com> #5941 authoritative over keen-bat-281 design-for-sign draft.
Regenerate DESIGN.md/ROADMAP.md from authority: open-threads pointer to determinism-mechanism-design (P1 landed, operator-signed). Supersedes keen-bat #5937 doc+ref without hand-editing generated artifacts. Co-authored-by: Cursor <cursoragent@cursor.com>
Contributor
Author
|
Verified cursor/composer-2.5 review on Spot-check against current tree:
Agree with APPROVE — P1 substrate modeling matches operator-signed shape; behavioral projection deferred to P5 as designed. |
Contributor
Author
|
Verified claude/claude-opus-4-7 review on
|
| layer | type | role |
|---|---|---|
dsl/std |
DeterminismAxis (Deterministic | NonDeterministic) |
single core authority — two arms only |
v2/std |
DeterminismClass (NonDeterministic { source: NonDetSource }) |
compiler refinement of the core — not a re-declaration |
behavioral (P5) |
OperationDeterminism (+ EventuallyConsistent sibling) |
service extension projecting from core |
The design explicitly rejects collapsing these into one type today (§2.0: EventuallyConsistent has no image in DeterminismClass; merging would nickname distributed semantics into compiler leak atoms). Adding source: NonDetSource to DeterminismAxis would violate the signed shape — coarseness is the point at the std layer; provenance lives in DeterminismClass.
Zero consumers at P1 is phased per §8 roadmap row P1 ("DeterminismAxis core in dsl/std + extend v2.std.determinism…"). P5 wires OperationDeterminism projection from DeterminismFact; until then the inert core is shape-sign, same pattern cursor noted for P1 phasing.
Not a parallel-authority fork — it's Vendor<Hardware>-style core + domain refinement (design §2.0 analog). The consolidation trigger is P5 behavioral projection + P4 #3468 infer bundle, documented in the signed design doc (operator-ratified authority for this PR).
non_det_source_eq / determinism_class_eq — acknowledged
Agree these are predicate/walker mechanical triggers, explicitly rostered in NON_FOLD_RESIDUE_ROSTER (non_fold_residue_project.rs:124-125) with kernel marker naming P4/#3468 dissolution — checked, not missed.
Contributor
Author
|
|
briansrls
added a commit
that referenced
this pull request
Jun 29, 2026
PR #5941 only added determinism eq sites; the three commit_workflow exceptions must stay on the roster until gunbc#5364 dissolves them. Co-authored-by: Cursor <cursoragent@cursor.com>
4 tasks done
briansrls
added a commit
that referenced
this pull request
Jun 29, 2026
…esses. Re-harvest closed #5941 onto fresh main (post-#5937 design merge): - v2.std.determinism primitive roster + determinism_compose algebra + witness data - dsl/std/determinism.dag DeterminismAxis authority - determinism_contract_test.dag (7 fail-closed witnesses) - design doc reconciled as MODIFY on #5937; DESIGN.md + ROADMAP.md regen'd - non_fold_residue roster: determinism_class_eq + non_det_source_eq Co-authored-by: Cursor <cursoragent@cursor.com>
briansrls
added a commit
that referenced
this pull request
Jun 29, 2026
…esses. (#5944) Re-harvest closed #5941 onto fresh main (post-#5937 design merge): - v2.std.determinism primitive roster + determinism_compose algebra + witness data - dsl/std/determinism.dag DeterminismAxis authority - determinism_contract_test.dag (7 fail-closed witnesses) - design doc reconciled as MODIFY on #5937; DESIGN.md + ROADMAP.md regen'd - non_fold_residue roster: determinism_class_eq + non_det_source_eq Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Cursor <cursoragent@cursor.com>
3 tasks
briansrls
added a commit
that referenced
this pull request
Jun 29, 2026
The three gates use different (source_roots, binary) tuples (verified): - DslCompileCleanGate: gunbc compile dsl+src/v2 - EmitDeterminismGate: gunbc compile dsl only (x2 oracle pair) - RegenVerifyGate: regen_stage0 --verify (different binary) No artifact sharing is possible across gates in the current set. M2's value is forward-proofing (future gates declaring the same tuple reuse the RunnableCompile node; duplicates caught by lens), not present-day savings. Displacement table corrected: M1 saves ~35s (resolve memo), M2 saves up to 1x compile when oracle pair collapses after #5941 closes the non-determinism gap. Also: M1 dissolution trigger corrected (M1 is orthogonal to M2, not subsumed by it). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
briansrls
added a commit
that referenced
this pull request
Jun 30, 2026
* Design sketch: floor shared-computation memoization (no implementation) Documents the root (double-paid full-tree compile: 4× subprocess + 2× in-process resolve_entry_graph), two fix axes (M1 within-walk resolve memo + M2 RunnableCompile artifact node), and the four operator decisions needed before any implementation lands. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix stale §1 framing of EmitDeterminismGate (oracle, not redundant) The original text said the x2 diff was redundant if compile is content-addressed. Corrected: the gate is a load-bearing oracle for known-live non-determinism; content-addressing assumes determinism not proves it. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix M2: each gate uses distinct compile tuple; no artifact sharing today The three gates use different (source_roots, binary) tuples (verified): - DslCompileCleanGate: gunbc compile dsl+src/v2 - EmitDeterminismGate: gunbc compile dsl only (x2 oracle pair) - RegenVerifyGate: regen_stage0 --verify (different binary) No artifact sharing is possible across gates in the current set. M2's value is forward-proofing (future gates declaring the same tuple reuse the RunnableCompile node; duplicates caught by lens), not present-day savings. Displacement table corrected: M1 saves ~35s (resolve memo), M2 saves up to 1x compile when oracle pair collapses after #5941 closes the non-determinism gap. Also: M1 dissolution trigger corrected (M1 is orthogonal to M2, not subsumed by it). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix §6 displacement table and §7 collapse claim Table row said '×2' (oracle pair survives) but the text below correctly said 4→3 (three distinct tuples remain after oracle pair 2→1). Fix to '×3' throughout. §7 'collapse to 1× total compile' was impossible: even after oracle pair collapses, DslCompileClean + EmitDeterminism×1 + RegenVerify are three distinct-tuple operations and all remain necessary. Corrected to '4→3×'. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix §6 displacement table: baseline 4× resolves (not 2×), M1 saves ~105s (not ~35s) Round 4 reviewer finding: Axis B actually has 4 resolve_entry_graph calls per run (Batch 1 DslCompileClean, Batch 2 SharedClaims group, serialized RegenVerify batch, serialized EmitDeterminism batch) — not 2×. M1 eliminates 3 redundant calls → ~105s saved, not ~35s. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix stale ~37s in §6 Total row — M1 saves ~105s not ~35s cursor review finding: Total compile cost row had ~37s in both M1 columns, inconsistent with the corrected ~105s in the M1 granular table row and prose. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Add link to floor-shared-compute-memoization sketch in DESIGN.md open threads Fixes doc_graph_has_no_orphan_docs CI failure: the new docs/plans file was not reachable from DESIGN.md. Added li to open_threads_blocks() and regenerated DESIGN.md via main_wet. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
briansrls
added a commit
that referenced
this pull request
Jun 30, 2026
) * Design sketch: floor shared-computation memoization (no implementation) Documents the root (double-paid full-tree compile: 4× subprocess + 2× in-process resolve_entry_graph), two fix axes (M1 within-walk resolve memo + M2 RunnableCompile artifact node), and the four operator decisions needed before any implementation lands. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix stale §1 framing of EmitDeterminismGate (oracle, not redundant) The original text said the x2 diff was redundant if compile is content-addressed. Corrected: the gate is a load-bearing oracle for known-live non-determinism; content-addressing assumes determinism not proves it. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix M2: each gate uses distinct compile tuple; no artifact sharing today The three gates use different (source_roots, binary) tuples (verified): - DslCompileCleanGate: gunbc compile dsl+src/v2 - EmitDeterminismGate: gunbc compile dsl only (x2 oracle pair) - RegenVerifyGate: regen_stage0 --verify (different binary) No artifact sharing is possible across gates in the current set. M2's value is forward-proofing (future gates declaring the same tuple reuse the RunnableCompile node; duplicates caught by lens), not present-day savings. Displacement table corrected: M1 saves ~35s (resolve memo), M2 saves up to 1x compile when oracle pair collapses after #5941 closes the non-determinism gap. Also: M1 dissolution trigger corrected (M1 is orthogonal to M2, not subsumed by it). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix §6 displacement table and §7 collapse claim Table row said '×2' (oracle pair survives) but the text below correctly said 4→3 (three distinct tuples remain after oracle pair 2→1). Fix to '×3' throughout. §7 'collapse to 1× total compile' was impossible: even after oracle pair collapses, DslCompileClean + EmitDeterminism×1 + RegenVerify are three distinct-tuple operations and all remain necessary. Corrected to '4→3×'. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * Fix §6 displacement table: baseline 4× resolves (not 2×), M1 saves ~105s (not ~35s) Round 4 reviewer finding: Axis B actually has 4 resolve_entry_graph calls per run (Batch 1 DslCompileClean, Batch 2 SharedClaims group, serialized RegenVerify batch, serialized EmitDeterminism batch) — not 2×. M1 eliminates 3 redundant calls → ~105s saved, not ~35s. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix stale ~37s in §6 Total row — M1 saves ~105s not ~35s cursor review finding: Total compile cost row had ~37s in both M1 columns, inconsistent with the corrected ~105s in the M1 granular table row and prose. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Add link to floor-shared-compute-memoization sketch in DESIGN.md open threads Fixes doc_graph_has_no_orphan_docs CI failure: the new docs/plans file was not reachable from DESIGN.md. Added li to open_threads_blocks() and regenerated DESIGN.md via main_wet. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * M1 within-walk resolve memo: fix Runnable pattern exhaustiveness Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * M1 resolve scope: add discriminating witnesses to ci_floor_plan GREEN: all heavy runnables in the plan declare ResolveScopeShared. RED: a heavy profile with ResolveScopeIsolated fails the lens. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * fix: OR use_walk_memo on SharedClaims merge; document source_roots invariant When group_batch_units merges a ResolveScopeShared SingleClaim into an existing SharedClaims unit that was created from an Isolated claim first, the group would silently drop the Shared flag and run via run_shared_entry_claims instead of the cross-batch memo path. Fix by ORing use_walk_memo on merge so the memo path wins whenever any member of the group declares ResolveScopeShared. Also document the memo-key invariant: source_roots is constant per run_walk call, so keying the memo by entry alone is safe today; notes what to change if that assumption ever widens. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: gate workspace_root import under #[cfg(test)] to clear unused-import clippy error Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: replace vec! with array literal in warm==cold test to silence clippy Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * §5 construction: drop resolve_scope — derive memo from heavy_whole_tree_resolve heavy_whole_tree_resolve is the single authority (§3). The memo decision follows by law: a whole-tree resolve is deterministic per source_roots, so heavy⟹memoize is always true. Two independent axes with one combination forbidden was validation-where-construction-was-available (DESIGN §5). Removes: type ResolveScope, resolve_scope field on RunnableResourceProfile, resolve_scope_eq, runnable_profile_resolve_scope_valid, and the three schedule walkers (runnable_resolve_scope_valid, batch_all_resolve_scope_valid, schedule_list_all_resolve_scope_valid, schedule_all_runnables_valid_resolve_scope). These dissolve — HeavyIsolated is now unwritable by construction, so the validator and witnesses testing it are unnecessary. claim_executor.rs derives use_walk_memo from profile.heavy_whole_tree_resolve directly (single read site, no gate list). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * reword stale ResolveScope vocab in two comments Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * WIP: DESIGN SKETCH ONLY (no code; returns to stern-moth-225 + operator before * fix: promote memo-eligible thread_units on entry cache hit; sync regen Two changes: 1. executor: at partition time, also move SharedClaims to memo_units if their entry is already in walk_memo (populated by a prior batch's heavy resolve). This eliminates cold re-resolves of the same entry in later batches even when the gate's own profile is non-heavy — achieving the documented 4×→1× resolve count per walk. 2. std_realization_schedule.rs: sync regen'd output (field shorthand vs explicit form in runnable_resource_profile constructor). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * add resolve-count oracle: memo fires resolve_entry_graph once per entry per walk memo_deduplicates_resolve_count: first call assert resolve_nanos > 0 (fresh resolve fires), second call for same entry asserts resolve_nanos == 0 (cache hit, resolve_entry_graph does NOT fire). Goes RED if the memo is bypassed. Discriminating witness for the 4x->1x dedup claim (DESIGN §2). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Brian Searls <briansrls@gunb.ai> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
briansrls
pushed a commit
that referenced
this pull request
Jul 4, 2026
…rator session 2026-07-04) Names the target architecture for the resolve problem measured during the CI dig-out: request-major private universes -> cheap edge graph + reverse-reachable minimal set + forward once-per-node evaluation (construction, not cache lookup), windowed-frontier memory (interface vs body artifacts), isolation-as-purity, and the full v1/v2 change surface. Staged S1 (union resolve, in-process, ships in the current PR line) / S2a (module nodes + Merkle keys) / S2b (persistence, gated on determinism #5941) / S3 (shared store) so it is not a 10-PR arc. Cross-linked from the M1/M2 memoization doc it supersedes (that doc's own M1 dissolution trigger anticipated exactly this design). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf
briansrls
added a commit
that referenced
this pull request
Jul 4, 2026
* CI witness opt-in inversion: zero witnesses by default, enrollment by declared roster row Operator decision 2026-07-04: the tree-wide glob roster (543 files / ~1,551 test fns, growing quadratically - a witness is definable between any two pipeline segments) pushed both CI jobs to the 90-min timeout: maximum cost, zero delivered signal, every merge effectively fail-open. Existence no longer enrolls a witness in CI. - CiSpec.witness_entries: the opt-in roster, projected from CommitWitnessClaim rows on the GithubActionsCiJob surface of commit_gate_roster (the existing enrollment authority - no new mechanism). Empty today. - ci_floor_plan: an empty roster puts NO witness-corpus node in the plan (structural omission, not an empty glob - the executor's empty-corpus fail-closed diagnostic stays meaningful); a non-empty roster becomes explicit-entries discovery batches (scan_dirs/discovery_scope_dirs empty, so the glob path is unreachable from CI), partitioned corpus vs execution by path prefix, still compile-gated and heavy-serialized. The dedicated grounding whole-tree node dissolved into the roster. - claim_executor: when a plan carries no DiscoveryBatch, run the naming- hygiene tree walk (test fn outside *_test.dag, __ basenames) once, fail-closed - a witness must stay nameable even when not enrolled. - Witness tests inverted: live plan asserts discovery tracks the roster (zero today); the enrollment machinery is proven on synthetic rosters (presence, partition, gating after compile, heavy serialization, exclusion edges) with red discrimination. - Rust lane, same inversion: the two whole-tree tests (func_env whole-tree ptr-count, whole-corpus semantic oracle) are now #[ignore] with written reasons per the #5427 discipline; run via -- --ignored. - Found and fixed two files the v1 interpreter grammar could not parse (top-level // comments): affected_set_floor_runner.dag - which made resolve_floor_runner_context fail and silently disabled the affected-set skip, falling back to full corpus on every floor run - and phase_profile_proof_plan.dag. Comments converted to data-string markers. - Local tree-wide discovery unchanged (claim_batch --roster-from-discovery). - Dissolve-on: affected-set selection + floor memoization make per-PR selection-by-affectedness affordable; enrollment returns to discovery shrunk by the affected set. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Enroll the first opt-in CI witness: variant_owner_expected_type (single pure-fn canary) Per operator intent: exactly one witness in CI for now, to observe the floor's behavior around a known-tiny payload. The canary is dag/test/claim/variant_owner_expected_type_test.dag :: expected_type_picks_variant_owner_not_alpha_order - a single test fn over self-contained pure functions (no host intrinsics, one std.logic import; measured locally: 3ms resolve / 2-module closure, ~0ms eval), so whatever time the CI witness lane now costs is the floor's fixed overhead, not the witness. Also: run the witness naming-hygiene walk unconditionally in claim_executor's plan path, BEFORE plan evaluation - explicit-entries discovery batches skip the glob scan that used to carry the placement checks, and pre-plan ordering makes a naming violation the cheapest possible failure. Verified by execution: red probe (planted test fn in a non-_test.dag file) fails closed with the placement diagnostic in seconds; green probe on the live tree passes the walk and proceeds to plan eval; the enrolled canary runs green via claim_batch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fix clippy absurd_extreme_comparisons on the emitted-Rust error ratchet (pre-existing on main; surfaces on any .rs-touching PR) The ratchet constant deliberately sits at its minimum (0); scoped allow with justification keeps raising the ceiling a one-constant edit instead of hardcoding == 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Eliminate every degradation arm in floor selection: declared machinery fails immediately Operator decision: 'fail closed' here means DON'T PROCEED, never 'silently run more'. skip_unaffected_node_frontier: true is a declared capability, so every input it needs is required - a failure is a loud typed error: - git diff observation failed -> error (was: run full corpus). The FloorGitDiffOutcome degradation enum is deleted; an absent diff is no longer a representable input state to selection. To run without selection, declare skip_unaffected_node_frontier: false - an explicit bit, not a silent state. - diff->declaration attribution failed -> error (was: run full corpus) - affected-set runner unresolvable -> error naming the declared entry (was: run full corpus - the arm that hid the broken runner for weeks) - same per shard -> error through the shard join (was: run all rows in shard) - precompute_would_skip / per-witness would_skip / frontier-touch fns erroring -> error naming the witness (was: run anyway) - roster row file unreadable -> error (was: silently reclassify as host-scaffold) Rationale: the 'safe' superset fallback conflated broken artifacts with missing observations, hid the defect permanently (no consumer of the eprintln), and at corpus scale 'run more' inverted into fail-open-by-timeout. The runner provisioning, fetch step, and observers are all declared by this repo's own spec, so there is no 'environment' excuse category: a missing input is a bug in a declared step and stops the floor. cargo build green, clippy -D warnings green, fmt clean; floor-scoped lib unit tests and a live red/green probe through a discovery batch are running and any regression they surface lands as a follow-up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Floor prelude: resolve the plan entry once, stamp every phase with wall-clock marks The 'gunbc ci' step went silent for 35+ minutes after the release build on the self-hosted runner - all of it pre-witness prelude: naming walk, output-policy install, plan-entry resolve, interpreted scheduler eval (gunbc_ci_floor_batches), then a SECOND resolve of the same plan entry for spawn-width, then another interpreted eval. Nothing printed until all of it finished, so a 30-minute prelude was indistinguishable from a hang. - resolve_entry_graph(plan_entry) now runs once; the hermetic plan eval and the wet spawn-width eval share the resolved graph (the double-paid-compute trap from the floor memoization thread, paid at minutes per resolve). - claim_executor stamps [t+Xs] phase marks: hygiene walk, policy install, plan resolved, plan evaluated, width evaluated / walk starting. The floor log now itemizes its own prelude. eval_plan stays as a resolve+eval wrapper for the --perturb-check path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * CI job timeout: 90 -> 10 minutes (operator inversion of the timeout ratchet) The prior ratchet went 30 -> 50 -> 90, each raise buying headroom for the unfixed resolve cost; at 90 the queue backed up and runs still died at the cap having delivered zero signal. The budget is now a forcing function: the floor must fit it, and the [t+Xs] phase marks itemize exactly which phase does not. Jobs stay red-by-timeout until resolve memoization lands - an honest red, unlike a 90-minute lane that never completed. ci.yml hand-synced with the ci_workflow.dag authority (3 job sites); the generated-artifact drift gate byte-verifies the pairing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver graph-major design doc: once-per-node module evaluation (operator session 2026-07-04) Names the target architecture for the resolve problem measured during the CI dig-out: request-major private universes -> cheap edge graph + reverse-reachable minimal set + forward once-per-node evaluation (construction, not cache lookup), windowed-frontier memory (interface vs body artifacts), isolation-as-purity, and the full v1/v2 change surface. Staged S1 (union resolve, in-process, ships in the current PR line) / S2a (module nodes + Merkle keys) / S2b (persistence, gated on determinism #5941) / S3 (shared store) so it is not a 10-PR arc. Cross-linked from the M1/M2 memoization doc it supersedes (that doc's own M1 dissolution trigger anticipated exactly this design). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Hand-sync generated docs to their edited authorities (DESIGN.md CI bullet, ci-selection-vs-scheduling section 7) The interpreted main_wet regeneration was killed twice by container restarts (~2 CPU-hours each attempt, itself a receipt for the resolver-graph-major design). The generated outputs are hand-applied to byte-match what the emitters produce from the already-committed authorities (dag/gunbc/design_document.dag, dag/gunbc/plans/ci_selection_vs_scheduling.dag); the GeneratedArtifactDriftGate adjudicates exactness once the floor fits its 10-minute budget. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * claim_batch: honor GUNBC_FLOOR_PHASE_PROFILE (same install as claim_executor) Without it, claim_batch diagnostics cannot attribute time to resolve/typecheck/eval phases - a 20-minute silent resolve is uninterpretable, and the resolver investigation needs per-phase receipts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: phase verdict (typecheck-dominant) + one-scheduler-no-special-cases section; per-module typecheck attribution Instrumented, isolated measurement: the 11-module closure finishes parse+resolve+normalize in ~1.1s then sits in typecheck 13+ minutes - inference is ~99.9% of cold resolve cost. Reprioritizes the plan: pathology lane first (per-module [typecheck-attribution] lines now name any module over 2s), unification (module obligations under the existing executor) second, union-of-global-passes demoted to a ~1s-class cut. Adds the operator's formulation as design section 1b: resolution is the system's one special case - the Bazel-shaped executor schedules eleven coarse gates while the dominant work hides in one opaque recursive host call; module resolution IS dependency execution over the same substrate. Maps the external witness-frontier-cursor vocabulary onto existing authorities (Runnable/RunnableCompile, Realization content-hash key + EffectShape, executor batches, receipts) per the section-3 anti-nickname discipline, and adopts the four-terminal-state invariant. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * S1a: process-level resolve store - fixed-entry machinery resolves once per process First implementation increment of the resolver graph-major design (and of the stratification rule: resolution is declared shared work, not ambient per-consumer control flow). A thread-local store keyed by (source_roots, entry) now serves the floor runner context, the diff observer, the output policy, the group syntax, the wet-hermetic roster prefix, and the executor's plan entry - each was previously a private resolve, re-paid per consumer and per discovery batch within one process. Failure semantics unchanged: a miss resolves exactly as before, including the typed error path; the store never converts an error into a fallback. Receipt: process_resolve_store_dedupes_repeat_resolve - Rc identity on the second resolve proves zero recompute (0.00s, fixture tree under target/). Thread-local by design: resolved graphs are Rc-based (not Send); shard threads keep their own store rather than smuggling Rc across threads. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * typecheck attribution: start markers (profile-gated) name the pathological module Verdict from the instrumented run: all std modules in the closure - including std.algebra and std.realization_schedule themselves - typecheck in under 2s each; the call that runs 17+ minutes is typecheck_module on the 80-line CONSUMER module (test.claim.realization_schedule_witness). The inference pathology is at generic use sites, not definition sites. Next lane: the existing type-env lookup profiler (text_lookup_work_counter feature, reset_type_env_lookup_profile) pointed at that one module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Phase heartbeat: live type-env amplification counters; pathology named Heartbeats now carry env_flatten/env_builds/env_merges/env_rewires from the existing v1_compiler_infer_env atomics. First live capture on the pathological entry: env_flatten grows ~80k/sec unbounded (9.7M by t=120s) while builds=11, merges=37, rewires=0 stay flat. Root site: record_lit_variant_fields_from_visible_env (v1_compiler_infer.rs, generated from v1.compiler.infer) - every VARIANT LITERAL flattens the entire visible environment (recursive parent fold; the counter) and then scans every visible binding with expand_type_for_field_access per candidate to find the variant's owner. Cost = variant literals x |visible env| x expansion - definitions cheap, consumers explode. 04_env.dag:35 carries an invariant expecting flatten_visible_parent_recurses==0 on import chains; live count says 9.7M in 2 minutes. Fix shape (next block, dag source + regen): a variant-name -> owner index built once per env, and expected-type short-circuit before any global scan. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: operator ruling - constructor ambiguity is an ERROR; resolution follows the binding edge The ambiguity support (bare constructor names legal across coproducts, expected-type tie-break, whole-env owner scan as fallback) is itself the bug - accidental semantics, never intended. The DAG already provides the namespace: imports bind arm names to arm nodes; the parent edge names the owner; resolution is following the edge already held. Rules: bound-name resolution via parent edge; unbound constructor literal = typed error (today a silent undeclared-dependency fail-open via the scan); double binding = collision error at env build; patterns resolve via scrutinee type; expected-type owner-picking deleted. The pathological scan is removed, not memoized. Census: 5,526 arm names, 221 global collisions (uniqueness must be scoped to co-visibility); <=4,572 unbound arm-name uses upper bound, dominated by legal pattern positions. Control run killed at 60+ min still inside one typecheck_module call on the 80-line module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: operator sequencing - union resolve first as interim with explicit contract; flat name visibility refinement Union resolve leads the resolve lane with a recorded contract: minimum upper bound (resolve cost <= 1x union closure, receipt-enforced), successor must be maximally parallel (topo-antichain module typecheck, budget-tree width) and frontier-window efficient; it dissolves into S2a and must not grow features that delay that. Namespacing ruling refined per operator: visibility is FLAT - locals + direct imports only, never transitive - making collision detection per-file with zero graph traversal; types propagate by graph identity underneath. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf --------- Co-authored-by: Claude <noreply@anthropic.com>
briansrls
pushed a commit
that referenced
this pull request
Jul 4, 2026
…erface Reframe the post-S1 step per operator direction: the "edges carry no payload" pre-req is not a module-specific blocker to clear before S2a — it is an interface generalization of the executor itself, added once and adopted by both riders. - §1b: the missing primitive is the artifact-bearing job (RunnableCompile shape) carrying produce/consume keys into the content-hash store; the CI floor rides it with a degenerate unit artifact, module resolution with ModuleInterface. One executor, N payloads. Scheduling reused unchanged. - §5: v2.workflow.executor is generalized to a dataflow executor, not reused as-is; migration receipt is a byte-identical 11-gate plan. - §7 S2a: two ordered moves — (1) generalize the executor + migrate the CI floor (behavior-preserving, unit payload), (2) put module resolution on the same executor (ModuleInterface payload). Not gated on determinism (#5941). Three open decisions flagged: keyed store vs inline edge payload, CI gates stay ordering-only for the migration, and who computes the 5.5k-node schedule (.dag authority vs host-computed + .dag-witnessed). Work lands after S1 merges; this is the plan only. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G6J28X52KhGXVArk6gcPgd
briansrls
added a commit
that referenced
this pull request
Jul 4, 2026
… shards) (#6234) * Union resolve S1: one shared index per process (prelude + discovery + shards) The request-major resolver gave every entry a private universe: the executor prelude did 3+ cold resolves (plan entry, output policy, group syntax), the discovery corpus resolved the floor runner in a separate index from its rows, and each shard rebuilt both. The shared std/spec prefix re-typechecked once per consumer — the top time-waster (docs/plans/resolver-graph-major-design.md §0). S1 (host-layer only, no seed regen, no corpus migration) routes every consumer through ONE shared MultiEntryIndex so the union closure typechecks once per node: - resolve_entry_graph_shared now resolves against a thread-local shared index (process_shared_index) instead of a fresh per-call module index, so the prelude's fixed entries union. - run_discovery_corpus_with_options builds one index threaded through frontier attribution, the floor runner (now resolve_entry_with_index, not a private resolve), and every row. Each shard shares one index for its floor runner + rows; cross-shard sharing is the S2b/Arc frontier (Rc<ResolvedGraph> is !Send). The switch to the index path is proven behaviorally identical to the cold resolve_entry_graph by the existing resolve_typed_cache_equivalence_test (cached == cold in every resolve order). Receipts (all execute, each with a red control): - once-per-node counter: union computes < sum of private closures, and re-resolving an already-resolved entry computes ZERO new typechecks (union_resolve_typechecks_each_node_once). Enforces the minimum-upper-bound contract — resolve cost <= 1x union closure, not Nx. - byte-identity oracle: union-view outcome == private resolve, every order (union_view_result_equals_private_resolve_in_every_order). - collision honesty: the shared typed cache fails loud on a module name mapping to two declaring files (check_module_source_identity), never silent-serve (source_identity_flags_coresidence_collision_but_allows_reexport). Same wall build_module_index already raises tree-wide, now at the cache seam. Interim: dissolves into S2a's parallel module-obligation architecture; no feature growth (design doc §7/§8). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G6J28X52KhGXVArk6gcPgd * Plan: S2a as general-executor consolidation, pre-req dissolved as interface Reframe the post-S1 step per operator direction: the "edges carry no payload" pre-req is not a module-specific blocker to clear before S2a — it is an interface generalization of the executor itself, added once and adopted by both riders. - §1b: the missing primitive is the artifact-bearing job (RunnableCompile shape) carrying produce/consume keys into the content-hash store; the CI floor rides it with a degenerate unit artifact, module resolution with ModuleInterface. One executor, N payloads. Scheduling reused unchanged. - §5: v2.workflow.executor is generalized to a dataflow executor, not reused as-is; migration receipt is a byte-identical 11-gate plan. - §7 S2a: two ordered moves — (1) generalize the executor + migrate the CI floor (behavior-preserving, unit payload), (2) put module resolution on the same executor (ModuleInterface payload). Not gated on determinism (#5941). Three open decisions flagged: keyed store vs inline edge payload, CI gates stay ordering-only for the migration, and who computes the 5.5k-node schedule (.dag authority vs host-computed + .dag-witnessed). Work lands after S1 merges; this is the plan only. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G6J28X52KhGXVArk6gcPgd --------- Co-authored-by: Claude <noreply@anthropic.com>
briansrls
added a commit
that referenced
this pull request
Jul 4, 2026
…ll, scan deleted — the engine PR (#6235) * CI witness opt-in inversion: zero witnesses by default, enrollment by declared roster row Operator decision 2026-07-04: the tree-wide glob roster (543 files / ~1,551 test fns, growing quadratically - a witness is definable between any two pipeline segments) pushed both CI jobs to the 90-min timeout: maximum cost, zero delivered signal, every merge effectively fail-open. Existence no longer enrolls a witness in CI. - CiSpec.witness_entries: the opt-in roster, projected from CommitWitnessClaim rows on the GithubActionsCiJob surface of commit_gate_roster (the existing enrollment authority - no new mechanism). Empty today. - ci_floor_plan: an empty roster puts NO witness-corpus node in the plan (structural omission, not an empty glob - the executor's empty-corpus fail-closed diagnostic stays meaningful); a non-empty roster becomes explicit-entries discovery batches (scan_dirs/discovery_scope_dirs empty, so the glob path is unreachable from CI), partitioned corpus vs execution by path prefix, still compile-gated and heavy-serialized. The dedicated grounding whole-tree node dissolved into the roster. - claim_executor: when a plan carries no DiscoveryBatch, run the naming- hygiene tree walk (test fn outside *_test.dag, __ basenames) once, fail-closed - a witness must stay nameable even when not enrolled. - Witness tests inverted: live plan asserts discovery tracks the roster (zero today); the enrollment machinery is proven on synthetic rosters (presence, partition, gating after compile, heavy serialization, exclusion edges) with red discrimination. - Rust lane, same inversion: the two whole-tree tests (func_env whole-tree ptr-count, whole-corpus semantic oracle) are now #[ignore] with written reasons per the #5427 discipline; run via -- --ignored. - Found and fixed two files the v1 interpreter grammar could not parse (top-level // comments): affected_set_floor_runner.dag - which made resolve_floor_runner_context fail and silently disabled the affected-set skip, falling back to full corpus on every floor run - and phase_profile_proof_plan.dag. Comments converted to data-string markers. - Local tree-wide discovery unchanged (claim_batch --roster-from-discovery). - Dissolve-on: affected-set selection + floor memoization make per-PR selection-by-affectedness affordable; enrollment returns to discovery shrunk by the affected set. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Enroll the first opt-in CI witness: variant_owner_expected_type (single pure-fn canary) Per operator intent: exactly one witness in CI for now, to observe the floor's behavior around a known-tiny payload. The canary is dag/test/claim/variant_owner_expected_type_test.dag :: expected_type_picks_variant_owner_not_alpha_order - a single test fn over self-contained pure functions (no host intrinsics, one std.logic import; measured locally: 3ms resolve / 2-module closure, ~0ms eval), so whatever time the CI witness lane now costs is the floor's fixed overhead, not the witness. Also: run the witness naming-hygiene walk unconditionally in claim_executor's plan path, BEFORE plan evaluation - explicit-entries discovery batches skip the glob scan that used to carry the placement checks, and pre-plan ordering makes a naming violation the cheapest possible failure. Verified by execution: red probe (planted test fn in a non-_test.dag file) fails closed with the placement diagnostic in seconds; green probe on the live tree passes the walk and proceeds to plan eval; the enrolled canary runs green via claim_batch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fix clippy absurd_extreme_comparisons on the emitted-Rust error ratchet (pre-existing on main; surfaces on any .rs-touching PR) The ratchet constant deliberately sits at its minimum (0); scoped allow with justification keeps raising the ceiling a one-constant edit instead of hardcoding == 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Eliminate every degradation arm in floor selection: declared machinery fails immediately Operator decision: 'fail closed' here means DON'T PROCEED, never 'silently run more'. skip_unaffected_node_frontier: true is a declared capability, so every input it needs is required - a failure is a loud typed error: - git diff observation failed -> error (was: run full corpus). The FloorGitDiffOutcome degradation enum is deleted; an absent diff is no longer a representable input state to selection. To run without selection, declare skip_unaffected_node_frontier: false - an explicit bit, not a silent state. - diff->declaration attribution failed -> error (was: run full corpus) - affected-set runner unresolvable -> error naming the declared entry (was: run full corpus - the arm that hid the broken runner for weeks) - same per shard -> error through the shard join (was: run all rows in shard) - precompute_would_skip / per-witness would_skip / frontier-touch fns erroring -> error naming the witness (was: run anyway) - roster row file unreadable -> error (was: silently reclassify as host-scaffold) Rationale: the 'safe' superset fallback conflated broken artifacts with missing observations, hid the defect permanently (no consumer of the eprintln), and at corpus scale 'run more' inverted into fail-open-by-timeout. The runner provisioning, fetch step, and observers are all declared by this repo's own spec, so there is no 'environment' excuse category: a missing input is a bug in a declared step and stops the floor. cargo build green, clippy -D warnings green, fmt clean; floor-scoped lib unit tests and a live red/green probe through a discovery batch are running and any regression they surface lands as a follow-up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Floor prelude: resolve the plan entry once, stamp every phase with wall-clock marks The 'gunbc ci' step went silent for 35+ minutes after the release build on the self-hosted runner - all of it pre-witness prelude: naming walk, output-policy install, plan-entry resolve, interpreted scheduler eval (gunbc_ci_floor_batches), then a SECOND resolve of the same plan entry for spawn-width, then another interpreted eval. Nothing printed until all of it finished, so a 30-minute prelude was indistinguishable from a hang. - resolve_entry_graph(plan_entry) now runs once; the hermetic plan eval and the wet spawn-width eval share the resolved graph (the double-paid-compute trap from the floor memoization thread, paid at minutes per resolve). - claim_executor stamps [t+Xs] phase marks: hygiene walk, policy install, plan resolved, plan evaluated, width evaluated / walk starting. The floor log now itemizes its own prelude. eval_plan stays as a resolve+eval wrapper for the --perturb-check path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * CI job timeout: 90 -> 10 minutes (operator inversion of the timeout ratchet) The prior ratchet went 30 -> 50 -> 90, each raise buying headroom for the unfixed resolve cost; at 90 the queue backed up and runs still died at the cap having delivered zero signal. The budget is now a forcing function: the floor must fit it, and the [t+Xs] phase marks itemize exactly which phase does not. Jobs stay red-by-timeout until resolve memoization lands - an honest red, unlike a 90-minute lane that never completed. ci.yml hand-synced with the ci_workflow.dag authority (3 job sites); the generated-artifact drift gate byte-verifies the pairing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver graph-major design doc: once-per-node module evaluation (operator session 2026-07-04) Names the target architecture for the resolve problem measured during the CI dig-out: request-major private universes -> cheap edge graph + reverse-reachable minimal set + forward once-per-node evaluation (construction, not cache lookup), windowed-frontier memory (interface vs body artifacts), isolation-as-purity, and the full v1/v2 change surface. Staged S1 (union resolve, in-process, ships in the current PR line) / S2a (module nodes + Merkle keys) / S2b (persistence, gated on determinism #5941) / S3 (shared store) so it is not a 10-PR arc. Cross-linked from the M1/M2 memoization doc it supersedes (that doc's own M1 dissolution trigger anticipated exactly this design). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Hand-sync generated docs to their edited authorities (DESIGN.md CI bullet, ci-selection-vs-scheduling section 7) The interpreted main_wet regeneration was killed twice by container restarts (~2 CPU-hours each attempt, itself a receipt for the resolver-graph-major design). The generated outputs are hand-applied to byte-match what the emitters produce from the already-committed authorities (dag/gunbc/design_document.dag, dag/gunbc/plans/ci_selection_vs_scheduling.dag); the GeneratedArtifactDriftGate adjudicates exactness once the floor fits its 10-minute budget. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * claim_batch: honor GUNBC_FLOOR_PHASE_PROFILE (same install as claim_executor) Without it, claim_batch diagnostics cannot attribute time to resolve/typecheck/eval phases - a 20-minute silent resolve is uninterpretable, and the resolver investigation needs per-phase receipts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: phase verdict (typecheck-dominant) + one-scheduler-no-special-cases section; per-module typecheck attribution Instrumented, isolated measurement: the 11-module closure finishes parse+resolve+normalize in ~1.1s then sits in typecheck 13+ minutes - inference is ~99.9% of cold resolve cost. Reprioritizes the plan: pathology lane first (per-module [typecheck-attribution] lines now name any module over 2s), unification (module obligations under the existing executor) second, union-of-global-passes demoted to a ~1s-class cut. Adds the operator's formulation as design section 1b: resolution is the system's one special case - the Bazel-shaped executor schedules eleven coarse gates while the dominant work hides in one opaque recursive host call; module resolution IS dependency execution over the same substrate. Maps the external witness-frontier-cursor vocabulary onto existing authorities (Runnable/RunnableCompile, Realization content-hash key + EffectShape, executor batches, receipts) per the section-3 anti-nickname discipline, and adopts the four-terminal-state invariant. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * S1a: process-level resolve store - fixed-entry machinery resolves once per process First implementation increment of the resolver graph-major design (and of the stratification rule: resolution is declared shared work, not ambient per-consumer control flow). A thread-local store keyed by (source_roots, entry) now serves the floor runner context, the diff observer, the output policy, the group syntax, the wet-hermetic roster prefix, and the executor's plan entry - each was previously a private resolve, re-paid per consumer and per discovery batch within one process. Failure semantics unchanged: a miss resolves exactly as before, including the typed error path; the store never converts an error into a fallback. Receipt: process_resolve_store_dedupes_repeat_resolve - Rc identity on the second resolve proves zero recompute (0.00s, fixture tree under target/). Thread-local by design: resolved graphs are Rc-based (not Send); shard threads keep their own store rather than smuggling Rc across threads. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * typecheck attribution: start markers (profile-gated) name the pathological module Verdict from the instrumented run: all std modules in the closure - including std.algebra and std.realization_schedule themselves - typecheck in under 2s each; the call that runs 17+ minutes is typecheck_module on the 80-line CONSUMER module (test.claim.realization_schedule_witness). The inference pathology is at generic use sites, not definition sites. Next lane: the existing type-env lookup profiler (text_lookup_work_counter feature, reset_type_env_lookup_profile) pointed at that one module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Phase heartbeat: live type-env amplification counters; pathology named Heartbeats now carry env_flatten/env_builds/env_merges/env_rewires from the existing v1_compiler_infer_env atomics. First live capture on the pathological entry: env_flatten grows ~80k/sec unbounded (9.7M by t=120s) while builds=11, merges=37, rewires=0 stay flat. Root site: record_lit_variant_fields_from_visible_env (v1_compiler_infer.rs, generated from v1.compiler.infer) - every VARIANT LITERAL flattens the entire visible environment (recursive parent fold; the counter) and then scans every visible binding with expand_type_for_field_access per candidate to find the variant's owner. Cost = variant literals x |visible env| x expansion - definitions cheap, consumers explode. 04_env.dag:35 carries an invariant expecting flatten_visible_parent_recurses==0 on import chains; live count says 9.7M in 2 minutes. Fix shape (next block, dag source + regen): a variant-name -> owner index built once per env, and expected-type short-circuit before any global scan. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: operator ruling - constructor ambiguity is an ERROR; resolution follows the binding edge The ambiguity support (bare constructor names legal across coproducts, expected-type tie-break, whole-env owner scan as fallback) is itself the bug - accidental semantics, never intended. The DAG already provides the namespace: imports bind arm names to arm nodes; the parent edge names the owner; resolution is following the edge already held. Rules: bound-name resolution via parent edge; unbound constructor literal = typed error (today a silent undeclared-dependency fail-open via the scan); double binding = collision error at env build; patterns resolve via scrutinee type; expected-type owner-picking deleted. The pathological scan is removed, not memoized. Census: 5,526 arm names, 221 global collisions (uniqueness must be scoped to co-visibility); <=4,572 unbound arm-name uses upper bound, dominated by legal pattern positions. Control run killed at 60+ min still inside one typecheck_module call on the 80-line module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: operator sequencing - union resolve first as interim with explicit contract; flat name visibility refinement Union resolve leads the resolve lane with a recorded contract: minimum upper bound (resolve cost <= 1x union closure, receipt-enforced), successor must be maximally parallel (topo-antichain module typecheck, budget-tree width) and frontier-window efficient; it dissolves into S2a and must not grow features that delay that. Namespacing ruling refined per operator: visibility is FLAT - locals + direct imports only, never transitive - making collision detection per-file with zero graph traversal; types propagate by graph identity underneath. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Union-resolve follow-up: main-thread discovery rides the thread's shared index run_discovery_corpus_with_options now takes process_shared_index(source_roots) instead of building a private MultiEntryIndex, so when the executor prelude already resolved on this thread, discovery reuses its parse/typed caches (one union per thread, not per consumer). Shards keep their own index (Rc !Send, per-shard contract in resolver-graph-major-design §7 S1). Also: resolve_typed_cache_equivalence_test fixtures move from std::env::temp_dir() to the workspace target/ dir — build_module_path_index fails closed on out-of-workspace paths, so the /tmp fixture could never resolve in a workspace-guarded environment (union_resolve_receipts_test precedent). Gates run: cargo build -p v1-compiler; cargo test -p v1-compiler-tests resolve_typed_cache + union_resolve (green); source_identity guard (green). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: three-layer executor framing (scheduler / runner+store / content keys) + realization tower Supersedes the 'artifact-bearing executor' framing (operator, 2026-07-04: artifacting is a separate concern, opaque to scheduling). §1b/§5/§7 now state the split: the scheduler is payload-agnostic and unchanged; dataflow lives in a runner + node-keyed store (typed_module_cache is its embryo); S2b only swaps the store's key function to content hashes. The S2a open decisions are settled by the split (results in the store, jobs reference by node identity; gates stay unit-result; .dag-authority schedule). New §5b: the realization tower — Rust is the bootstrap row, not the floor; the intrinsic kernel is the declared HAND_MAINTAINED surface + registered builtins; realization-scoped decisions (Rc !Send) carry dissolve-ons. Post-fix timing receipts are placeholders by design, filled at the receipts commit of this PR. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Constructor-owner ruling (§1c): binding-edge resolution, per-file collision wall, scans deleted — dag sources Implements the operator ruling in the .dag authority (v1.compiler.infer / infer_resolve). This commit is step 1 of the seed two-step: the checked-in generated Rust still realizes the OLD semantics; the next commit hand-patches the seed functionally, and the regen commit replaces the hand-patch with this source's true emission (RegenVerifyGate byte-identity + bootstrap fixed point). - infer_record_lit: the unconditional whole-env scan-first rung is DELETED (record_lit_variant_fields_from_visible_env and both call sites). The ladder is now instantiate -> expected -> declaration/binding-edge only. - variant_owner_node: O(1) owner lookup — scope.locals binds arm name to its owning coproduct NODE; no name round-trip, owner need not be name-visible. - build_module_context: constructor namespace is FLAT — local coproduct arms + direct imports' contributions (is_all covered; specific enum or arm names covered; owner node carried on the binding). The transitive ancestry fold and the unique-owner-or-nothing import map are deleted. - Collision wall: one arm name -> two different owners in one file appends VariantCollision at env construction (insert_variant_owner_checked); the same declaration via two import paths is one owner, not a collision. Kernel coproduct arms keep low-priority prelude merge. - expected_type_override_enum (expected-type-as-owner-picker) DELETED; ExprVar variant bindings resolve by binding edge alone. - field_in_any_variant_named re-scoped to the resolved owner's arms. - Unbound constructor literal now reports UnresolvedType (was a generic InternalError message); no new diagnostic variants minted — VariantCollision and UnresolvedType already existed in 00_core. - collect_unit_variant_phantom_matches (04_resolve): flat one-level search (own + direct parents' str_bindings), no recursive ancestry flatten. - Dead code deleted: unique_imported_variant_owner, enum_parents_for_variant_in_items, is_imported_variant_owned_by, union_parent_variant_locals, fold_local_coproduct_variant_locals (kernel path inlined), variant_locals_from_items. Note: v2.std.determinism roster rows ^unique_imported_variant_owner / ^alpha_sorted_variant_fold are quoted symbols and still compile, but now name superseded fns — flagged for operator re-signing (#5941 roster). Baseline receipt (pre-fix, this base): realization_schedule_witness_test (81 lines, 2 imports) killed at 900s still inside typecheck_module, env_flatten 69.7M and linear (~78k/s) — target/baseline/single_module_pre_fix.log. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Constructor-owner ruling: hand-patched seed (two-step step 2) + diagnostic tests + payoff receipt Functional hand-patch of generated v1_compiler_infer.rs ONLY (201 insertions, one generated file — budget was ~200 across <=3): (a) build_module_context realizes the new flat, collision-walled construction (local arms + direct imports; owner node on the binding; VariantCollision on two owners for one name; kernel arms low-priority); (b) variant_owner_node helper; (c) record_lit_expected_fields owner via the binding node, scan fallback deleted; (d) infer_record_lit scan-first rung deleted; (e) ExprVar expected-type owner-picker deleted; (f) unbound constructor literal -> UnresolvedType. Deferred to regen (semantic gap re-censused with the true seed): field_in_any_variant_named re-scope, phantom-path flat re-scope, dead-fn deletions. This hand-patch is replaced wholesale by regen output in the convergence commit (precedent 3d89b2d/f3cc9f5). PAYOFF RECEIPT (the C4 stop/go gate): realization_schedule_witness_test.dag (81 lines, 2 imports) — pre-fix killed at the 900s cap still inside typecheck_module, env_flatten 69.7M and linear (~78k/s); patched seed: 1.14s TOTAL, env_flatten 0. Receipt logs: target/baseline/. New constructor_owner_ruling_test.rs (7 tests, green): intra-file collision, cross-import collision, unbound->UnresolvedType, enum-import/arm-import/ glob-import (brace-less form) construct green, re-export non-collision. The collision wall found live prey immediately (sweep work, next commits): dag/std/markup.dag Fragment.TextNode vs MarkupNode.TextNode (converted cross-owner at :183 — the expected-type pick in vivo); integer.dag DecimalDigit vs NonZeroDecimalDigit D1..D9; target_model.dag TargetReferenceLayer{,Wrapped}; plus bare-None uses that resolved via the transitive scan. Corpus-resolving tests stay red until the sweep lands — fix-then-sweep ordering, full-suite green re-established pre-regen. Also: dependency_pool_index whole-tree compile tests (release-binary corpus walks) join the 2026-07-04 opt-in inversion; they had escaped it by self-skipping on CI (no release binary) while running unbounded locally. Gates run: cargo build -p v1-compiler; cargo test -p v1-compiler-tests constructor_owner_ruling (7/7) + resolve_typed_cache + union_resolve + source_identity (green); payoff receipt above. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Constructor-owner ruling: re-export chains + witness rewrites + fixture hygiene (C4 follow-through) Re-export chains (dag authority + seed hand-patch, both sides): a module re-exports its specific-name imports (get_exported_names), so 'import proxy { B }' where proxy itself imports B is an explicit import under the ruling. New owner_of_exported_arm / exported_coproduct_item walk the acyclic import chain to the defining module's coproduct; build_imported_variants resolves every specific name through them (glob imports stay own-items-only, matching the export surface which excludes is_all pass-through). Receipt: all 4 self_gen8 re-export chain tests green. Hand-patch grows to ~280 lines / still 1 generated file — over the ~200 budget line, accepted because the patched seed is the sweep census tool and without chains it mis-censuses every re-export site in the tree. Old-semantics witnesses rewritten to witness the ruling: - variant_owner_disambiguation_test: expected-type-pick / local-shadow / per-site tests -> collision red-controls + a binding-edge emission green control (fixtures preserved). - type_env_scope_chain_test: local-shadows-imported -> collision red-control with sole-owner green control. - pipeline: ambiguous_variant_name_resolves_correctly deleted (redundant with constructor_owner_ruling_test collisions); duplicate_variant_names_across_enums_dont_collide -> rule-4 witness (pattern-position arms resolve via scrutinee with ZERO constructor bindings; emitted match still owner-qualified). Fixture hygiene: - 6 files moved off std::env::temp_dir() to workspace target/ (12 tests; build_module_path_index fails closed outside the workspace — the union_resolve_receipts_test precedent, same class as the C1 fix). - 9 pipeline wire-contract fixtures gain 'VariantNaming' in their std.serialization imports (SnakeCase & friends are its arms — the rule-(i) migration class, in fixture strings). Suite: 626 passed / 28 failed / 85 ignored in 167s. All 28 remaining failures are enumerated corpus-sweep work (integer.dag D1..D9, extdeps github Pending + openai UrlCitation collisions, dag/std markup TextNode, samsung.dag Gen3, target_model + src/v2 test-fixture bare-None sites) or the stale release binary (interp_dry_run) — zero unexplained. They re-green with the sweep commits, before regen. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Sweep wave 1 (regen-input closure): 3 census findings fixed Patched-seed census over src/v1 + its dag closure (89 sources, 21s — the same scope pre-fix was the 48-minute class) found exactly three violations: 1. dag/std/computation.dag: SizeBound.TreeSize collided with the imported RankingDimension.TreeSize in the module's own scope. Renamed the bound arm to SubtreeSize (the more precise fact — a tree-descent bound measures the shrinking subtree per step; rename fully contained in one file, SizeBound is imported by name nowhere else). RankingDimension keeps TreeSize (~12 sites in v1.complexity/std.induction). 2. src/v1/02_parse.dag constructed InternalError without importing it — previously resolved by the deleted whole-env scan (an undeclared- dependency fail-open, exactly the census class the ruling predicted). 3. src/v1/04_infer.dag used FieldNotFound without importing it (same class, in the compiler's own source). Generated-seed counterparts of the SubtreeSize rename land at the regen commit (all TreeSize references in stage0 are in generated files). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Sweep: markup TextNode collision + floor-closure unbound imports Fragment.TextNode -> FragmentText (dag/std/markup.dag): Fragment and MarkupNode both declared TextNode in one file, and markup_node_to_fragment converted one to the other cross-owner via the deleted expected-type pick. MarkupNode (the larger surface, ~15 consumer files) keeps TextNode; the 5 Fragment-side consumers follow the rename (std.markdown, extdeps languages/markdown, serializer witness, regime2 plan prose, medium-structure fixture). Floor-closure unbound sites from the first CI run of #6235 (exactly the census class - constructors resolved by the deleted scan without imports): fleet_posix_accounts gains ConstructionMechanism (SingleAuthority's owner); samsung + western_digital storage rows gain PcieGeneration + NvmeFormFactor (Gen3/Gen4/M2_2280 owners). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * CI canary swap + constructor-owner red-control suite The enrolled CI canary asserted the DELETED semantics (expected-type picks the variant owner), so it is swapped for github_merge_state_witness_test.dag :: witness_clean_and_blocked_are_distinct_variants - small, pure, and exercises imported-variant construction through the new binding-edge path (operator decision 3, 2026-07-04). The retired canary fixture becomes the collision RED control: new diagnostics_witness suite 'constructor_owner' (VariantCollision on the two-local-owners fixture; UnresolvedType on an unbound constructor literal; sole-owner green control), wired through the existing transport pair (tools.diagnostics_witness_transport + test.claim.diagnostics_test). variant_owner_expected_type_test.dag is deleted from the compiled tree - its fixture is now illegal by design and lives on inside the red control. Verified by execution: target/debug/diagnostics_witness constructor_owner green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Sweep round 1 (10 parallel agents): 51 collisions renamed + 67 unbound sites imported across src/v2 + dag closures The whole-tree census with the patched seed drove file-disjoint fix groups; policy per operator decisions: bigger owner keeps the arm name, smaller side renames along its distinguishing dimension; unbound constructor literals gain their owner-enum import from the defining module. Collision renames (owners in parens keep the shared names): - cache catalog twins: Catalog-side arms -> CatalogOpaqueStringKey / CatalogCasContentDigest / CatalogOpaqueStringEncoding / CatalogCasContentEncoding (std.cache_interface keeps); the catalog->std converter in extdeps/cache/cache.dag now reads per-side by scrutinee/produced type instead of the deleted expected-type pick; 5 cache witness suites green by execution. - integer digits: NonZeroDecimalDigit arms D1..D9 -> NonZeroD1..D9 (DecimalDigit keeps D0..D9); widen fn patterns renamed, constructed DecimalDigit results kept; subset-type modeling flagged as a §3 follow-up rather than done ad hoc here. - cpp_abi: singleton width enums' arms -> CppWidth8Fixed..CppWidth64Fixed (CppIntegerWidth keeps CppWidth8..64); per-model core-width arms -> CppILP32/LP64/LLP64/ILP64CoreIntegerWidthModel; anchor scalar -> SignedIntegerScalarAnchor (CppScalar keeps SignedIntegerScalar). - language width models (rust, ptx, kotlin, java, go, swift): int-width enums keep BitsN; float/complex/ordered-ring/etc. sides renamed along their dimension (FloatBitsN, ComplexBits64, ...). - effects + target_model: smaller-side renames per the same policy (WrappedReferenceLayerRc/Box on TargetReferenceLayerWrapped). Unbound-import fixes: 30+ files across src/v2 (manual/execution/lens test claims, std, compiler stages, program.dag, extdeps languages/ formats, workflow runner test) — each gains exactly the owner-enum import its constructor literals need. Bare-None sites resolved per expected type (enum arm import vs the lowercase optional keyword). VERIFICATION (executable gate): whole-tree census re-run with src/v2 entries — 290 error lines before, 3 after, and all 3 are deliberate red-fixture plants (parse-error + layering-scan plants, other walls' fixtures). Zero constructor-ruling findings remain in src/v2 scope. dag-root verification runs next alongside sweep round 2 (wave-2 findings: srv3 aggregate, same-file pairs, witness-test imports). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Sweep round 2 (4 parallel agents) + env-counter dissolution (unblocks regen) Round 2 fixes the dag-tree census findings: - srv3_os_install_diagnostic: 13 collision pairs resolved. The aggregate Srv3InstallDiagnostic KEEPS its arm names (larger use counts + owns the wire strings); sub-verdict copies renamed with the file's existing ...Verdict suffix convention (ReadyToBootVerdict, InstallerHungVerdict, OsInstalledVerdict, InconclusiveVerdict, ServePortAbsentVerdict, ServePortConflictVerdict, TokenMissingVerdict, IsoMissingVerdict, KvmWeakUntrusted) and OsInstallRuntimeDiagnosticVerdict's copies gain the Runtime prefix. All wire strings byte-identical. - Same-file pairs: ReviewSource.Llm -> LlmSource (ReviewProvider keeps); ReviewEvent.Pending -> PendingEvent (ReviewState keeps the upstream-cited PENDING; not a wire event value upstream, so no cited string breaks; pipeline.rs wire-contract needle updated); OpenAI annotation UrlCitation -> ChatUrlCitation / ResponsesUrlCitation (both endpoints cited, zero uses, no wire-contract rows exist for these enums); ConvergeVerdict.Converged -> VerdictConverged (std Reconciliation keeps; fleet_show_effective_read's if-branch type mismatch was this collision's downstream shadow); workflow/types ArtifactKind.Design -> DesignDoc and OutcomeStatus.TerminalFailed -> OutcomeTerminalFailed (IssueLifecycleStage keeps both). - Witness-test + gunbc/extdeps import batches: reviewer_source, hardware selection, transport fidelity (S/L/Xs size arms), pep440 (Ordering arms), uri path tokens, jedec/bmc/storage grounding rows, ci_render, gunbhub_serve, sk_hynix, transports file/rest/shell - each gains its owner-enum import from the defining module. Env-counter dissolution (04_env.dag + 04_infer.dag): the 2026-07-04 pathology instrument's record_* statement calls and stub fns are deleted - the pathology they measured is dead (900s-capped -> 1.14s), and they were also the ONLY corpus instances of a statement-then-expression block the Rust emitter renders without a separator. That emitter deficiency is why regen has been silently broken since B1/B2 (whose commits hand-edited generated files): the first regen probe failed parsing emitted v1_compiler_infer_env.rs at exactly that construct. Getter stubs stay for phase_profile API compatibility. Emitter statement-separator fix is ledgered with the emit-stage work (regen probe: Rust emit completes in ~2 min, so C8 is time-feasible). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * regen_stage0: register phase_profile.rs as hand-maintained The 2026-07-04 heartbeat module was added to stage0/src without a registry entry, so --emit-fresh assembled a crate whose lib declares mod phase_profile but carries no file (probe failure after the counter dissolution unblocked emission). Its header carries its own dissolution trigger (realization_measurement_loop Phase 0). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Rust emitter: shared_types dominates needs_box_wrapping (kills Box<Rc<T>> double-wrap) A shared (Rc-rendered) type never needs Box - Rc is already the cycle-breaking indirection. The old ordering short-circuited recursive->Box before the shared check, which was benign only while emit-time recursive_types stayed module-local; after B1's build_type_env rework propagated imported recursive types, every Rc field of Node/TypeEnv/PositiveDescentAmount/... double-wrapped, and regen output stopped matching the committed seed ABI (596 of the 770 build errors on the first regen attempt). This is one of the deficiencies that accumulated while RegenVerifyGate was red-by-timeout: B1/B2 co-landed hand-edited generated files because regen could not reproduce them. Two-step: dag authority (05_emit_rust.dag) + functional hand-patch of the generated emitter (v1_compiler_emit_rust.rs), replaced by regen's true output in the convergence commit. Receipt: re-regen after the reorder emits ZERO double-boxed fields (the only remaining Box<Rc grep hit is this fix's own marker string); committed Box<ByteSize>-style non-shared boxing (std_realization_schedule) preserved exactly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fix emitter-fix commit: restore hand-patched needs_box_wrapping (prior commit captured regen output) The previous commit's v1_compiler_emit_rust.rs accidentally snapshotted the REGENERATED emitter (regen write had overwritten the hand-patch before the commit): the regenerated emitter carries the deref-side boxing asymmetry and does not build. Restored the committed seed's emitter with the shared_types-dominates reorder applied as intended. Receipt: cargo build --workspace green on this commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Resolver design: emitter restoration ledger (§7b) — regen was semantically broken on main The constructor-ruling PR's regen attempt discovered RegenVerifyGate's red-by-timeout had been masking semantic regen breakage (B1/B2 co-landed hand-edited generated files). Three deficiencies fixed in this PR (statement-separator emission via counter dissolution; needs_box_wrapping shared-dominates reorder; phase_profile registry); three ledgered with receipts for the restoration PR (deref-side boxing asymmetry ~800 errs, alias-brand rendering ~250, misc ~80). Seed stays the functionally-verified hand-patched realization until restoration lands. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Diagnostics: locate error-typed nodes reaching evaluation Two instrumentation edits for the fail-open seam where an inference-side error node reaches the interpreter without a blocking diagnostic: - v1_interpreter (hand-maintained): ExprError eval errors now carry the node's span (file:start-end). - infer field-access cascade (seed hand-patch, replaced at regen): names the accessed field, the error-typed base, and the module. Receipt: the parse_table_memo failure went from 'error type cascade' to 'error type cascade (field class on error-typed base tok in v2.compiler.parse) at src/v2/compiler/02_parse.dag:31281-31282' — the open investigation for the 7 remaining interpreted-parse suite reds. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fix two pre-existing silent compiler seams (interpreted-parse suite reds) + walls Both located via span-instrumented errors, a 14-fixture minimal-repro ladder, and pristine-main worktree A/B (both reproduce identically on main; both classes were unverifiable there through the pre-fix resolve timeouts). 1. Interpreter kernel-optional unwrap (hand-maintained v1_interpreter.rs): kernel optionals at runtime are raw-value-or-Null, and match_pattern's raw-payload unwrap guards (Present+Optional, Holds+Witness) sat BELOW the kind-specific arms - Value::Record/List/Str/Int matched their kind arm first and returned None from inside it, so 'match xs |> first { Present { value: t } => ... }' failed non-exhaustive on any record element (Variant payloads had an inlined fix; everything else fell through). Guards hoisted above the kind arms; Variant payloads excluded so the existing Variant-arm logic stays authoritative. 2. Kernel-prelude generic shadowing (dag authority 04_resolve.dag + seed two-step in v1_compiler_infer_resolve.rs): for multi-import modules the ancestry cache merges the kernel as OVERLAY, so a user generic coproduct NAMED Optional (v2.std.collection) resolved to the paramless kernel Optional at its consumers; is_user_generic_use_site went false, Optional<T> signatures were never expanded/stamped, lookup_variant_in_type returned Blocked with ZERO diagnostics, and pattern bindings went error-typed silently - the 'error type cascade' at eval (import-count-dependent because single-import modules skip the kernel overlay). Fix is surgical: a use site WITH type arguments whose primary lookup lands on a paramless decl retries the DIRECT import parents' str_bindings for a parameterized decl (the params filter naturally excludes the kernel parent). The wholesale precedence flip (kernel-as-base) was tried and reverted - it broke kernel function dispatch through an unmapped path; ledgered for the restoration PR. Receipts: parse_table_grammar_memo_multi_file_ingest_parses PASS (was 'error type cascade at 02_parse.dag:31281'); all 11 repro fixtures PASS; new permanent walls in kernel_shadow_seams_test.rs (record-payload unwrap; imported-Optional shadowing). Located-diagnostics kept: ExprError eval errors now carry spans; the infer cascade names field/base/module. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Post-engine-PR roadmap: deferred ledger + compiler algorithm audit plan Part A: the twelve items PR #6235 deliberately deferred, weighted - emitter restoration -> regen convergence (§7b receipts), emit-stage cost (the next resolver-class pathology, >20min/89 modules measured), kernel-prelude shadowing root rule (wholesale flip blocked on the unmapped kernel-fn-dispatch path), Value::Null split, flat-visibility re-census decision, C9 receipts, determinism roster re-signing, binder find-first hardening, diagnostics dedup, witness re-enrollment dissolve-on, coverage-by-illusion census, resolver S2a/S2b/S3. Part B: the algorithm audit method - every stage factored as units x work-per-unit (the resolver failed both axes at once), inventory -> verdict -> measure -> fix by displaced cost -> budget witness per stage. Survey table lands from the eight-stage parallel inventory in flight. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Brace else-match in resolve_generic_use_decl (v1 grammar requirement) The parents-retry fold used 'else match' without braces, which the v1 grammar rejects. Claim-path closures never parse src/v1, so only resolve_expr_types_retraversal_guard_test caught it ("parse failed for src/v1/04_resolve.dag"). Braced; guard test green. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fix interp_dry_run hermetic fixture: workspace target/, not /tmp The failure was misdiagnosed as a stale release binary. Real cause: the hermetic witness fixture was written to std::env::temp_dir(), and build_module_path_index fail-closes on module paths outside the workspace — same seam as the 7 test files already migrated; this one was missed. Fixture now lives under workspace target/ like the rest. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Roadmap: record local full-suite green receipt (656/0/85) Corrects the A.6 stale-binary diagnosis: the last two suite reds were the resolve_generic_use_decl else-match brace and the interp_dry_run /tmp fixture path, both fixed in the prior two commits. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Algorithm audit: 8-stage survey results + ranked fix order Appends the survey synthesis to the post-engine-PR roadmap: 9 PATHOLOGICAL findings ranked with anchors (dag_collect Rc-accumulator O(M^2) as the measured >20-min emit suspect; token-stream skip|>first O(M^2) parse; whole-corpus rewire scan per name; per-entry closure fixpoint regression vs the in-file worklist BFS; request-major imported-variant binding; interpreter env chain O(d); v2 translate facts/serialize quadratics; dead v2 packrat memo; v2 closure-chain namespace), a SUSPICIOUS tier, six root-cause clusters, and a fix order priced by displaced cost. Raw per-stage catalogs land as their own receipt doc. Four top claims spot-verified against the live tree. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * rustfmt: format hand-edited interpreter guards + new test files CI's cargo fmt --check flagged the hoisted match_pattern guard block in v1_interpreter.rs, the fixture-dir join chain in kernel_shadow_seams_test.rs, and the out-of-alphabetical-order mod declaration in tests/lib.rs (rustfmt reorders modules). cargo fmt --all applied; fmt --check, clippy -D warnings, and the seam tests verified green locally. Also enabled the repo pre-push hook path in this clone. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf * Fold hand-edited doc content into dag authorities; main_wet regen green C9 item: main_wet regen + drift gate. Three findings, all fixed at the authority layer, not the projection: - PR #6186 added §11 (folded deltas) to the generated bounded-input-cost-envelope-scheduling.md without updating its dag authority — a parallel-representation slip the drift gate caught on this branch's first hooked push. §11 is now authored in dag/gunbc/plans/bounded_input_cost_envelope_scheduling.dag and the projection regenerates identically to the signed content. - regime2-shared-emission-fold.md regenerates with FragmentText (this branch's markup collision rename), ROADMAP.md with normalized tail. - Doc-graph orphans: linked the post-engine deferred-ledger/audit roadmap from the CI lane's receipts prose (survey receipt doc is reachable through it), and linked docs/plans/host-converge-inventory.md (orphaned since #6177 — no inbound link anywhere on main) from the converge-lane prose. Receipts: PASS main_wet, PASS run_generated_artifact_drift_gate_body, PASS doc_graph_has_no_orphan_docs, PASS doc_graph_has_no_dangling_links. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DcouXUHJmm4NPqwDyvnDKf --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
docs/plans/determinism-mechanism-design.md(FLAG 1 locked: bundle determinism into T-13 — lens family expansion: lens/parallelism.dag + lens/effect.dag + lens/ownership.dag + lens/idempotency.dag (one mechanism, projection-only per Practice 11; lens-over-substrate enforced; reads InferredTree + dependency-graph projection; produces Witness<dimension>; canonical reference = T-12 PR #3468InferredFactsblock).dsl/std/determinism.dagcoreDeterminismAxis(two-arm universal authority).v2.std.determinismwith closed primitive roster (incl. Ground emit-path variant-owner determinism in .dag authority #5913 construction rows), left-biaseddeterminism_compose,DeterminismFactcarrier, and inline witness data.determinism_contract_test.dagwith §5 MVP discriminating claims (pure-add GREEN, hash-map-keys-iter RED, compose algebra).Test plan
gunbc run --claim-runon all 5determinism_contract_testwitnesses → greenclaim_executorresolve onv2.std.determinism+ test module → clean