test(build): pin arbitrary module nesting and DFS embed order

[Chemaclass]

Summary

build.sh already supports module directories nested to any depth, but nothing pins that
behaviour. The two properties the distributable depends on are untested, so a future change
to build::process_file could break the released binary without any suite going red.

Current behaviour (verified, not assumed)

build::process_file is a depth-first walk of source statements, not a directory walk:
it emits a file's body (shebang stripped via tail -n +2), then recurses into every
^source line, dedupes on the repo-relative path (#923), and the final artifact has all
source lines stripped.

Verified on main (7eecc45) with a throwaway three-level module:

src/depthtest.sh -> src/depthtest/mid.sh -> src/depthtest/deep/leaf.sh

Result: all three bodies embedded, in DFS order, 0 source lines remaining in the output,
and the depth-3 function present in the single-file binary. So the flat→module migration
(#924, #925, and the rest) needs no build work — but that is currently an undocumented,
untested property.

Why it is worth pinning

Two regressions have already shipped through this exact seam:

• Watch subcommand fails with "bashunit::watch::run: command not found" (src/watch.sh missing from release build) #735 — src/watch.sh was missing from the release build entirely.
• fix(build): key embed markers on repo-relative paths so src/ can hold module dirs #923 — the embed dedupe was keyed on basename, which both hid a genuine double-embed and
  would have collided src/parallel.sh with src/runner/parallel.sh the moment a module
  directory appeared.

Both were build-graph bugs invisible to the normal suite, because the normal suite runs from
src/, not from the built artifact.

Proposal

Add regression tests to tests/unit/build_test.sh (which already holds
test_module_aggregators_hold_only_source_lines_and_comments), building into a temp dir from
a fixture tree rather than touching real src/:

1. Arbitrary nesting is embedded. A three-level fixture
   (aggregator → mid aggregator → leaf) ends up fully embedded, leaf included.
2. DFS order is preserved. A file's own body precedes the bodies of the files it sources —
   this is the property that makes the "aggregators hold only source lines" rule load-bearing,
   and nothing currently asserts it.
3. No source lines survive in the artifact.
4. Same basename in two modules does not collide. src/<a>/parallel.sh and
   src/<b>/parallel.sh both embed — the exact fix(build): key embed markers on repo-relative paths so src/ can hold module dirs #923 regression, currently only implicitly
   covered because src/parallel.sh and src/runner/parallel.sh happen to coexist.

Constraints

• Bash 3.0+ floor.
• Fixtures under tests/**/fixtures/ must not end in *test.sh — the runner would collect
  them as test files.
• Build into a temp dir ($(bashunit::temp_dir)); never write to bin/ from a test.
• tests/unit/build_test.sh sources build.sh to exercise its functions directly — follow
  that existing pattern rather than shelling out to a full build where possible, and keep the
  test off the fork-budget hot paths.

Acceptance criteria

• Tests added to tests/unit/build_test.sh covering the four properties above
• Each fails for the right reason when the corresponding behaviour is broken (verify by
  temporarily inverting the logic, then reverting)
• ./bashunit tests/ · ./bashunit --parallel tests/ ·
  ./bashunit --parallel --simple --strict tests/ green
• make sa && make lint green
• bash build.sh bin -v still prints ✅ Build verified ✅
• Bash 3.0+ compatible
• No CHANGELOG entry — internal, no user-visible behaviour change

Do not

• Do not change build.sh behaviour — this issue only pins what it already does
• Do not create tests/unit/build/; make test globs one level only
• Do not run shfmt -w; make lint is the format gate

[Chemaclass]

Done in #933 (merged).