peel_alias_once_for_field_access drops resolve_node diagnostics: type the probe-miss vs refusal split, audit diagnostic-count consumers

[briansrls]

Summary

peel_alias_once_for_field_access was a speculative alias-expansion probe, but it projected only NodeResolveResult.resolved. That made a legitimate zero-diagnostic probe miss indistinguishable from a resolver refusal and silently removed diagnostics before inference could carry them.

This change returns and recursively accumulates the existing NodeResolveResult carrier, then appends its diagnostics in both judgment-producing consumers: ExprFieldAccess and record-literal alias field discovery. It mints no parallel outcome vocabulary and does not change resolver policy. Variant-membership probes and emitter shape reads remain metadata-only and explicitly project .resolved.

Impact audit

Two projections dropped these diagnostics before CompileResult.diagnostics: direct field-access expansion and record-literal alias field discovery. All downstream readers of that one carrier therefore undercounted this class:

• compile-clean hard/advisory predicates and counts (compile_clean_diagnostic_is_hard, compile_clean_diagnostic_is_advisory, and their vector/count consumers);
• compile-clean diagnostic histogram identity (compile_clean_diagnostic_histogram_key);
• the substrate-visible diagnostic census (compile_dag_diagnostic_census), at exact (diagnostic_class, subject_name, blocking) identity grain.

Zero finding: the unresolved-method frontier occurrence budget does not consume UnresolvedType identity, so its counts are unaffected.

This is why the namespace N1 census must be rerun after this fix: its previous kernel excluded peel-suppressed diagnostics from both inference consumers by construction. Completing the record-literal path raises the standalone whole-tree advisory count by five (5,880 → 5,885), so the provisional revealed delta was necessarily low and the freeze census must pin the post-merge compiler.

The standalone whole-tree control remains at zero blocking errors because this PR changes propagation, not the type-reference admission policy: unadmitted pool hits are still accepted until the N1 refusal arm in draft PR #7622 lands. The newly observable hard UnresolvedType population therefore appears under the combined peel fix + N1 arm, which is why #7630 must land before that branch reruns and freezes its census.

Test plan

• CTRL_BUILD_BYPASS_SHIMS=1 cargo test -p v1-compiler peel_alias_fixpoint_termination --lib -- --nocapture — PASS, 2 tests. The carrier controls prove two UnresolvedType refusals survive recursive peel, a legitimate non-alias Int probe stays at zero, and the exact pre-projection identity set is a subset (retired_count == 0). A second test compiles an in-memory module through compile_sources and pins distinct field-access and record-literal alias refusal multiplicities in CompileResult.diagnostics. Detached RED controls remove each append independently: field identity count falls 4→2 and record identity count falls 3→1, so either consumer dropping diagnostics fails the witness.
• CTRL_BUILD_BYPASS_SHIMS=1 cargo run -p v1-compiler --bin regen_stage0 -- --verify — PASS, regen_divergence_count=0.
• rebuilt target/debug/gunbc, then ctrl-build --local -- target/debug/gunbc compile --output-dir /tmp/witty-badger-200-record-lit-compile.FYLUMF --source-root dag --source-root src/v2 --target dag — PASS across 1,702 sources: 0 blocking errors, 5,885 advisory diagnostics (five newly reported through record-literal alias inference).
• cargo fmt --all -- --check / pre-commit and pre-push formatting gates — PASS.

Attestation

• The title describes the change.
• The branch contains the dashboard WIP checkpoint and the intentional completion commit; no foreign commits or cherry-picks.
• No secrets, credentials, or large binaries are included.

CI red is inherited from main — session_dashboard.dag is unenrolled in the extdeps scope rosters (fix pending in #7640/#7633); it is the sole failing check, and all approvals are in.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[gunbai-bot]

Addressed review 46723 on current head 3745446277:

• Added the DESIGN §7 seed-retention receipt beside the host-thread/timeout witness. It names the concrete v1-test-migration prerequisite and the ROADMAP zero-hand-written-Rust row as its dissolution trigger.
• Kept the .resolved projection in record_lit_alias_struct_fields intentionally. That helper (and field_in_any_variant_named) is a metadata-only shape probe used to choose a declaration-field view; it does not own a diagnostic ledger or admit an executable artifact. Making that speculative lookup emit would convert a probe into a new record-literal refusal path—the false-red direction this change is required to exclude. The judgment-producing ExprFieldAccess path is the consumer that now carries the complete NodeResolveResult diagnostics into InferResult. The .dag authority note now states this caller split explicitly.

Focused witness remains green (quiet=0, refusal=2, retired=0) and regen_stage0 --verify reports zero divergence.

— sent from witty-badger-200

[gunbai-bot]

Review 46744 (and the first finding in review 46723) is addressed at 5d8b74895b.

record_lit_alias_struct_expansion now returns the existing NodeResolveResult; infer_record_lit takes fields from .resolved and appends .diagnostics to the final InferResult. Refused expansion can no longer silently supply record-literal field expectations. The branch preserves the prior short-circuit evaluation order, so unused later probes are not evaluated eagerly.

The rebuilt whole-tree compile is clean across 1,702 sources with 0 blocking and 5,885 advisory diagnostics. That is five additional diagnostics over the field-access-only head, confirming record-literal alias inference was a second undercount source. The PR impact statement now records this and the requirement that the freeze census pin the post-merge compiler.

The hand-Rust receipt requested by review 46723 remains present and names v1-test-migration plus the zero-hand-written-Rust ROADMAP row. Focused witness and fixed-point verification are green.

— sent from witty-badger-200

[gunbai-bot]

Review 46772 is addressed at 85db3ba9fc with a production-ledger witness.

The new test compiles an in-memory .dag module through compile_sources, exercising both changed inference consumers and asserting their distinct refusal identities in CompileResult.diagnostics. It passes with the fix (FieldAlias=4, RecordAlias=3). Detached RED controls removed each append independently: the field-access count fell 4→2, and the record-literal count fell 3→1; each removal fails its own assertion. The existing direct carrier test still covers quiet probe miss, reveal, and retirement independently.

cargo fmt --all -- --check and both focused tests pass. The PR body records the compile-ledger witness and both RED measurements.

— sent from witty-badger-200