[Ch4s3]

Summary

• forge test compiles all discovered test files together but picks only ONE as the compile entry (alphabetically first — forge/lib/cmd_test.ml's run_files).
• Reachability pruning (fix(resolver): resolve modules by declared name; stop silently dropping test modules #106) keeps a non-entry sibling test module's own declarations unconditionally, since DDescribe/DTest count as "global effect". But it never fed that sibling's own textual references back into the reachability BFS — the seed set was only the entry file's source plus explicit-import sources.
• Net effect: a plain library module referenced only by a non-entry sibling test file (never by the entry file itself) is silently pruned from the merged program. This surfaces as a misleading Unknown module Foo. Did you mean Bar? at typecheck time — not a pruning diagnostic — for a module that plainly exists and is plainly used.
• This isn't an edge case: any multi-file test suite where the alphabetically-first test file doesn't happen to textually mention every lib module some other test file needs will hit this. Found it while upgrading sigil (well, a fork under development) to current main: test/markdown_test.march sorts before test/sigil_test.march and never mentions Page/Build, which sigil_test.march uses extensively — so those two lib modules vanished from the compiled test binary.

Fix

Seed the reachability queue with the source text of every discovered module that's kept unconditionally via has_global_effect_decl, in addition to the entry file and explicit-import sources.

Verification

• New regression test test_lib_module_used_only_by_sibling_test_is_not_pruned, confirmed to fail with only the fix reverted (Expected true, Received false) and pass with it.
• run_compiler: 615/615
• run_eval: 256/256
• Manually verified against sigil's real 3-file test suite (markdown_test.march, properties_test.march, sigil_test.march, ~1800 combined lines): the Unknown module Page/Build errors are gone after this fix.
• Rebased onto current main (past fix(join-points): never float a let that reads the case scrutinee #114 and fix(refinecheck): a stdlib import must not withdraw the List.length alias #115) — the List.length alias bug I'd originally bundled here landed independently as fix(refinecheck): a stdlib import must not withdraw the List.length alias #115, so this PR is back to just the resolver fix.

Test plan

• run_compiler full suite green
• run_eval full suite green
• New test fails on revert, passes with fix