Namespace Blocker-1 (part 1): reference-derived module-edge producer + inert-lens reach repoint

[gunbai-bot]

Summary

Namespace Blocker-1 (part 1): the native reference-derived module-edge producer the namespace terminal step ("deps via container.member") needs, plus repointing the inert-lens hygiene reach to it.

The module dependency graph is import-parse-derived (extract_import_paths). After the corpus-wide import strip it goes empty, so the inert-lens hygiene reach BFS flags ~20 v2.lens.* modules as inert — the strip PRs' documented ~44s CI blocker.

reference_resolution_facts parses each import-less file once through the real front-end (tokenize + parse — no substring scan), collects qualified-name chains + bare identifier use sites, and resolves them to the declaring module via a global exported-name index:

• same-module names resolve locally (no edge) — keeps a ubiquitous fixture data like live_tree_disposition (declared in ~670 files) from fanning every referrer out to every declarer;
• cross-module names disambiguate by containment proximity (nearest module-path prefix — Bool → the nearest v2.std.logic);
• edges carry a confidence tag (Qualified | UniqueBare | AmbiguousBare); the inert-lens reach reads the strict set (dropping AmbiguousBare) so an over-connected graph can never silently clear a truly-inert lens (DESIGN §5 — no fail-open hygiene).

The reach path_imports is now import ∪ reference per file: an import-bearing file uses its exact import edges (unchanged); a stripped file is carried by its reference edges. Monotone reach ⇒ zero regression by construction (the reached set only grows, so the inert set can only shrink).

What is NOT here (staged part 2)

The module-graph loader closure stays import-derived. A reference-derived loader changes every witness's load set tree-wide and surfaces latent issues — import-less-but-referencing std files, witnesses that need src/v1 in their pool, the pre-existing fleet_converge Srv3 red, and homonyms the bright-cat lane must qualify — so it is staged after those land (see the note in build_module_graph_facts_live_uncached).

Test plan

• Build: cargo build --release — clean, no warnings.
• Inert-lens hygiene on a locally-stripped src/v2/test tree (claim_batch --roster-from-discovery --source-root dag --source-root src/v2 --scan-dir dag/test/claim --scan-dir src/v2/test/claim/manual): no inert-lens error (was 20 inert without the repoint).
• Producer's only live consumer is the inert-lens reach; monotone union ⇒ unstripped inert count unchanged (0).

[gunbai-bot]

CI red is inherited Blocker-2 (main-head compile-clean), not this diff.

The failing check is dag_compile_clean_gate_passes (returned Bool(false)) in floor batch 1. I reproduced the whole-tree compile-clean locally and it fails on exactly one pre-existing defect:

non-exhaustive match: missing variant(s) Srv3InstallReconcileObserve, DurableApprovalGrantRecord,
DurableApprovalGrantRead, Srv3InstallMediaFetch, Srv3SeededInstallMediaRemaster,
Srv3SeededInstallMediaToolchainEnsure, Srv3SolConsoleCapture, Srv3BmcwebSessionLogin,
Srv3BmcwebTokenExtract, Srv3NbdProxyServe, Srv3BootOnceReadBackSleep, Srv3ReceiptEmit
(dag/gunbc/fleet_converge_cli.dag:2974-2979)

This is a semantic merge-skew: shell→dag #6586 (P5a) added 12 Srv3*/DurableApproval* HostEffect variants, and the converge_cli match arm in dag/gunbc/fleet_converge_cli.dag was not updated to cover them. It reproduces on origin/main with imports fully restored, so it is entirely disjoint from this PR — tracked as Blocker-2 (dag/gunbc / shell→dag lane), a separate dispatch.

This PR's diff is a single file (src/v1/stage0/src/cli_run.rs, +510) that adds a seed-side reference-derived module-edge producer consumed only by the inert-lens hygiene reach (import ∪ reference-strict). The module-graph loader (build_module_graph_facts_live) is deliberately left import-derived — the loader repoint (part-2) is parked until the homonym-qualification lanes land. Nothing here touches fleet_converge_cli.dag or the HostEffect match, so it cannot cause or fix the compile-clean red. CI here will go green once Blocker-2 lands on main.

— sent from fierce-ram-489

[gunbai-bot]

Thanks — on the fixture-cleanup nit: the test does clean up on success. reference_edges_derived_from_bare_refs_local_and_import_aware calls std::fs::remove_dir_all(&root) as its last statement (cli_run.rs:12776), after all three asserts, and also pre-cleans at the start (12733). The only path that leaves the fixture behind is an assert panic (a test failure), where unwinding skips the trailing teardown — and even then it's a pid-namespaced dir under gitignored target/ (gunbc-refedge-core-<pid>), so it never touches version control and is self-limiting.

I'm not pushing a commit for it: (a) it's cosmetic and only on the failure path, as you noted non-blocking; (b) a new commit resets the SHA, which would invalidate the freshly-refreshed 2/2 approvals and re-run CI on this merge-critical PR. If we want panic-safe teardown, the natural place is a Drop guard, and it can fold in when this scaffold is revised — the whole reference_edge_producer_tests module dissolves at the named trigger (symbol_index_fill) anyway.

— sent from fierce-ram-489