Whole-tree typecheck advisory carrier: ResolveTypecheckGate::DiagnosticsCollector (north-star #5760/#5772)

[gunbai-bot]

What

The sole carrier authority for the north-star whole-tree typecheck work (driving down the ~779 orphan-module diagnostics #5760 parse-resilience unmasked). Builds the collect-and-continue gate + its first advisory consumer. Advisory only, non-blocking — promotion to a blocking gate is the operator's call once the count is driven to ~0 (DESIGN §5/§6; do not flip a floor gate from closure-scope to whole-tree-scope unilaterally).

Per the anti-fork pin: this PR is the single home for the WholeTreeCtx/whole_tree_resolved_ctx signature change; sunny-fox's #5774 drops its own carrier changes and consumes DiagnosticsCollector after this merges.

Carrier (cli_run.rs)

• ResolveTypecheckGate::DiagnosticsCollector — no diagnostic is blocking; the full resolve/typecheck diagnostic set is returned (collect-and-continue), vs Strict (fail-closed) and DiscoveryCorpusAdvisory.
• resolved_graph_from_sources returns diagnostics as a 3rd tuple element; WholeTreeCtx carries a resolve_diagnostics field; whole_tree_resolved_ctx takes the gate as a parameter (was hardcoded Strict). Existing 3-arg callers (the wiring-liveness whole-tree bin + its enum test) pass Strict — behavior preserved.

Consumer (whole_tree_typecheck_advisory.rs)

Thin advisory bin: resolves every module under the source roots in one pass under DiagnosticsCollector, prints one TSV line per diagnostic (module_name<TAB>file<TAB>message) for downstream classification, always exits 0. This is what makes resolve_diagnostics live (read by execution, not just typechecked).

Discriminating test (green by execution)

whole_tree_diagnostics_collector_test.rs over a 2-module fixture (one clean, one with a dangling call):

• Strict fails closed; DiagnosticsCollector resolves both modules and surfaces the diagnostic naming wtd.broken + the missing symbol — the gate delta proven, not asserted.
• a clean-only subtree collects zero diagnostics (no-fabrication control). Flip the dangling call to a real one and both gates go green — the red-witness control.

Corrected measurement: 179 → 8 (de-fabricated exclude)

The earlier "179 diagnostics / 22 modules" count was 96% a measurement artifact of the advisory bin's own default exclude, not corpus state. The default was a blanket ["test/fixture/"] — a path substring match that drops the top-level shared fixtures real test modules import (derivable_coercion_task_id, rung_3_4_common, rung_5_6_common, task_manager_demo, …). Excluding an imported module manufactures a phantom unresolved-import in every importer plus its cascade — the derivable_coercion=49, find_witness=29, nat_semiring rung clusters were all this.

This PR narrows the default to scanner/lens fixture INPUT subdirectories only (read by path, not importable: layering_scan/, transport_script_scan/, realization_vocab_scan/, medium_structure_scan/, extdeps_external_authority/, floor_skip/, program_assembly/) via a single-authority helper default_advisory_fixture_excludes. The corrected whole-tree count is 8 diagnostics / 6 modules.

The narrowing is justified by principle, not hand-tuned to a number, and pinned closed so it cannot silently re-broaden: the new test default_advisory_fixture_excludes_are_scanner_only asserts no entry is the blanket root, every entry is a test/fixture/<dir>/ subdir, imported shared fixtures are NOT excluded, and scanner inputs ARE (DESIGN §5 — a fabricating exclude is a fail-open measurement; pin it closed).

The remaining 8 are not corpus rot: a single root no field 'Env'/'Which' on type 'shell' — modules call shell.Env.Get / shell.Which.Check, and extdeps.shell does define service shell.Env + service shell.Which (shipped tools that resolve in-closure). This is a genuine whole-tree resolver service-member gap (QN-collision / service-resolution family), addressed in a separate follow-up PR (stage0 resolver, skew-prone — deliberately not folded here).

CI note

The only CI red expected is the inherited floor pair, which clears on #5821 (merged). This PR also folds a one-line fix for an inherited fleet-red (#5777 left a dead duplicate let extdeps_root that reds clippy --all-targets -- -D warnings on every .rs PR) so this PR's clippy gate is a true green signal — separate labeled commit.

🤖 Generated with Claude Code

[gunbai-bot]

CI red @ d531b63 is infra-OOM, not a carrier regression — no fix to push.

Both failing jobs (ci, rust_tests) die at the same point with the same signature:

sccache: Compiler killed by signal 9
error: could not compile `v1-compiler` (lib)  (exit status: 254)

That is rustc itself SIGKILL'd while compiling v1-compiler lib — at CARGO_BUILD_JOBS=1, on the BUILD step, before any test or witness runs. There is no FAIL [batch N] marker and no error[...] compile diagnostic — the discriminators of a real red. The carrier diff is 3 small files (advisory-exclude helper + bin wiring + a pin test); it compiles clean and passes fmt + clippy (-D warnings) + the 3 carrier tests locally. Host memory pressure is OOM-killing the rustc process, which is the host-converge tentpole tracked at the fleet level, not anything in this PR.

Action: HOLD the frozen HEAD (d531b63). No re-push, no re-run — re-running just re-rolls the same OOM dice. This clears when the host-converge cap applies. Already approved (claude, zero findings); the remaining gates are a 2nd distinct provider + host-converge, both operator-owned.

— sent from smart-wolf-187