[Ch4s3]

Three resolver bugs found while moving forgepm onto current main. Each is independently verified and has a regression test that fails with only its own fix reverted.

1. A module was importable only at its conventional path

import Depot.Query failed with Module `Depot.Query` not found (looked for `depot/query.march`) for depot's own modules, which live at lib/api/depot_query.march. forge check passed and forge build did not, so the two disagreed about what resolves.

Nine Depot* names sat in stdlib_module_names, whose documented job is suppressing "not found" for modules that ship in stdlib/. None of them do — depot is a third-party package. While import_refs yielded only a path's first segment that bogus entry was accidentally load-bearing: use Depot.Query was skipped as stdlib and auto-discovery supplied the module. Once 7849f0d made candidates dotted, the stdlib filter dropped the bare Depot and the surviving Depot.Query became a hard error. (URI was listed too, but stdlib/uri.march declares mod Uri, so the casing mismatch meant a plain import Uri reported "not found".)

Deleting those entries alone leaves the real gap: find_file resolved only by the conventional dotted→path mapping, while auto-discovery indexes by the mod declaration — so a module whose file doesn't follow the convention was reachable by reference but not by explicit import. find_file now falls back to scanning for the file that declares the name, only when the conventional path misses, so a genuinely-absent module keeps its precise looked for <path> diagnostic.

This is not Depot-specific: mod Config.Test hit the identical failure, and Config genuinely does ship in stdlib.

2. forge test silently ran a fraction of the suite

forge test compiles ONE test file as the entry and relies on auto-discovery for the siblings. Reachability pruning (48cf926) keeps a discovered module only when the entry references it or it carries a global-effect decl — and nothing references a test module by name, which is what a test module is. Every sibling was dropped, collect_tests never saw its bodies, and the run reported a PASS over whatever survived.

forgepm went from 544 tests to 21, exit 0. Silent test loss with a green exit is the worst available failure mode.

DDescribe/DTest/DSetup/DSetupAll now count as global-effect. DDescribe is the one that matters — a test file's top-level decls are describe blocks, and adding only DTest changes nothing.

3. Checking order discarded the search path

Fixing (2) then produced 552 expected () but got Unit errors. Phase 2 sorted discovered modules by NAME (dot-count, then alphabetical), discarding the order the search path was given in. forge deliberately puts lib/ before test/ on MARCH_LIB_PATH, and its own comment explains why: a lib function still bound to its pass-1 Mono placeholder gets pinned by whichever call site is checked first. Forgepm.Test.* and Forgepm.Web.* have equal dot-counts, so Test sorted ahead of Web and pinned every shared helper from a test call site. Search-path index is now the primary sort key; the dot-count rule still orders within a directory.

Verification

• run_compiler 607/607, run_eval 256/256
• Both new tests verified to fail with only their own fix reverted. The pruning test writes a real entry file — resolve_imports disables pruning outright when it cannot read source_file, so a synthetic path makes such a test vacuous (my first version passed either way)
• End to end: depot, bastion and forgepm all build against this. forgepm goes from 21 tests, 0 failures (unfixed main) to 544 tests, 0 failures, matching the pre-regression count on the old pin

Related

Needs march-language/depot#1 and march-language/bastion#5 to build those two against current main; neither is required for this PR.