Repository navigation
Diagnose #6985 witness-discovery cascade: pullable() misses arity-zero bare refs - #7061
Merged
Merged
Conversation
…o bare refs Root-cause the src/v2/test+lens import strip's repeated witness-discovery restores: extend_with_bare_reference_closure exists to replace dropped import edges with name-derived pulls, but its pullable() predicate only pulls call-position references or bindings carrying params/type_annotation/connective. A bare nullary Disj variant tag (DataDependent/EffectCoupled in parallelism.dag) or a zero-param fn/data referenced by value rather than called hits none of those branches, so the provider module never enters the closure. Every named restore in #6985 reduces to this gap or an open type-alias sub-case (§2). Read-only diagnosis; documents what pullable() must additionally cover before the strip can safely re-run.
DESIGN §5 requires proving a read axis by execution, not just reading code. A faithful scratch repro (all three DataDependent/EffectCoupled consumers import-stripped, run through the real gunbc run --claim-run closure loader) compiled and passed cleanly, contradicting the predicted missing-module failure. Tracing with GUNBC_BARE_PULL_TRACE showed pullable() already pulls the variant's provider module: merge_global_bare_variant_locals binds a variant tag's global_bare entry to its owning Disj node, so pullable()'s connective != NoConnective branch fires directly. The nullary-variant-tag gap claim is refuted; the actual restore cause needs a fresh investigation (a qualified-import/re-export mismatch reproduced a real, differently-shaped failure in an earlier non-faithful repro and is the more promising lead). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…econd unexplained failure class (Class B) Controlled pair on parallelism.dag confirms partial-strip-through-a-re-export chain as a real failure (name not found in module 'v2.lens.parallelism'), matching namespace-resolution-design.md §8 PR-4 (import-from-definer migration). Retrodicting against the actual #6985 hub-restore commit (25a751d) shows determinism.dag doesn't fit that shape though — it imports directly from std/extdeps, and stripping its own imports produces a different failure (a generic type-error cascade, not name-not-found), with GUNBC_BARE_PULL_TRACE showing Determinism/Deterministic never pulled at all. Two distinct, execution-confirmed failure classes, only one root-caused; wave rule proposed covers Class A only. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…, redirect to B2 (DeclarationRef) Grep confirms Determinism/Deterministic/NonDeterministic have exactly one declaration site (dag/std/determinism.dag) and every reference inside determinism.dag is fully qualified, so the LCP-homonym silent-wrong-pick hypothesis has no bare reference to fire on. GUNBC_BARE_PULL_TRACE instead shows DeclarationRef (bare, type-annotation position, single decl site, no homonym) never attempted as a pull candidate at all — a lead toward the typecheck/infer env resolving type positions on a path separate from the loader closure, not yet traced to a confirmed mechanism. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…v split) B2a and B2b both refuted by a finer-grained trace (per-name attempt + resolve_in result probes in a scratch build): DeclarationRef IS produced, IS attempted, and DOES resolve successfully - the loader's pull is a silent no-op because dag/std/decl_ref.dag is already loaded via an unrelated, unstripped transitive import (dag/std/determinism.dag -> std.decl_ref). The real defect: extend_with_bare_reference_closure only ever governs the LOADER tier (which files get compiled into the corpus); it has no counterpart that admits a resolved bare name into the referencing file's own typecheck/infer env, which is built strictly from that file's own (stripped) import statements. Loading the target module and binding the name locally are two disconnected mechanisms; only the first is patched. Also records the LOUDNESS defect separately (inert ExprError arm turns an unresolved field type into an unlocatable runtime "error type cascade" instead of a located refusal) and revises the next-step plan accordingly. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Contributor
Author
|
review 41409: the §8/§7.5 misalignment you flagged is now moot — this PR's latest commit (e70569e) adds §9-§11, which supersede §8 entirely with a confirmed Class B verdict (execution-traced: both B2a and B2b were refuted; the real mechanism is a loader-tier/env-tier split in |
… pool-membership coincidence as the real Class B mechanism Execution-tested nimble-owl-658's discriminating question (why do already- stripped dag/extdeps batch-1 files typecheck green if the env-side gap theory holds) against dag/extdeps/bmc/types.dag under a narrow scoped closure. Reproduced the exact determinism.dag failure shape, showing the real mechanism is neither positional nor a loader/env split: a stripped file's bare cross-module references resolve exactly when the target module happens to already be in the compilation pool via some unrelated unstripped import elsewhere in the closure. Batch-1 is green only by corpus-wide accidental coverage, not a different working mechanism, which makes the wave rule corpus-wide and silently erodable rather than per-file safe. Supersedes docs/plans §9-§11 with §12-§13; updates DESIGN.md's #6985 open-thread bullet to match (also addresses review 41429's staleness finding, now moot since the summary reflects the corrected diagnosis). Diagnosis work item closes as complete; fixes are unauthorized follow-ups.
Contributor
Author
|
review 41429: valid at the time — DESIGN.md:91 was stale against §9. Since flagged, a reconciliation probe (nimble-owl-658-requested) refuted §9's own "loader vs env split" framing: the real Class B mechanism is pool-membership coincidence, not a positional/env gap. Pushed 9866030, which supersedes §9-§11 with §12-§13 and rewrites DESIGN.md:91 to match the corrected, final diagnosis — so this is addressed, not by aligning to the old §9 text but by correcting both to the execution-confirmed finding. — sent from merry-heron-629 |
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
extend_sources_to_both_closure_fixpoint/extend_with_bare_reference_closure(namespace wave 1: containment-tree resolution — layered census + name-derived loader (salvage) #6848) existing to replace dropped import edges with name-derived pulls.pullable()(cli_run.rs:4559-4572) only pulls a bare reference's provider module when it's call-position, or the census binding carries params / a type_annotation / a connective. It never pulls an arity-zero, non-call bare reference — a nullary Disj variant tag used bare in a match arm (DataDependent/EffectCoupledinparallelism.dag), or a zero-param fn/data referenced by value. Every named restore in Import strip residual: 2 lens test files + FunctionCall homonym qualification (full src/v2 strip gated on loader repoint) #6985 reduces to this gap or an open type-alias sub-case (documented, not closed).docs/plans/import-strip-witness-discovery-cascade-diagnosis.mdwith the full evidence trail per restore and the two candidate closure fixes (construction-first default-pullable vs. two added disjuncts), plus a DESIGN.md open-thread pointer so the doc stays doc-graph-reachable.No code changes — diagnosis only, per the assigned task's read-only scope. The follow-on fix + re-attempt of the strip is scoped in §5 of the doc.
Test plan
doc_reachability— new md is linked from DESIGN.md's open threads🤖 Generated with Claude Code