…e v0.0.147

Closes #628.

## Root cause

`vera/codegen/modules.py`'s cross-module harvest loop populated
`_fn_sigs` via `setdefault` from each imported temp generator
but never touched `_fn_ret_type_exprs`.  That registry — added
in #614 (for `IndexExpr` element-type inference on
`Array<T>`-returning calls) and re-used in #602 (for `String`
interpolation segments) — stores each user-declared FnDecl's
full Vera return-type AST.  Inference walkers look up
`_fn_ret_type_exprs.get(call.name)` to figure out whether a
function call returns `Array<T>` or `String`.

For cross-module callees that lookup returned `None`, the
walkers fell through to the silent-skip path that #602 / #614
had already closed for in-module callees.  Two failure shapes:

1. `make_arr(())[0]` where `make_arr: @Unit -> @array<Int>` is
   defined in a different module: `_infer_index_element_type`
   returns None → enclosing function dropped via `[E602]` → no
   exported `main` to run.

2. `IO.print("\(make_str(()))\n")` where `make_str: @Unit ->
   @String` is cross-module: interpolation segment falls
   through to `to_string(...)` silent wrapper, the `i32_pair`
   `make_str` return trips `expected i64, found i32` at WASM
   validation.

## Fix

One-block addition in `vera/codegen/modules.py`'s cross-module
harvest loop, alongside the existing `_fn_sigs.setdefault`:

```python
for fn_name, ret_te in temp._fn_ret_type_exprs.items():
    self._fn_ret_type_exprs.setdefault(fn_name, ret_te)
```

Same shape as the `_fn_sigs` harvest above.  `setdefault` so
first-seen wins — no extra collision detection needed since
the existing `_fn_sigs` collision loop above already rejects
name clashes before reaching this code.

## Regression tests

Two new tests added to
`tests/test_codegen_modules.py::TestCrossModuleCodegen`:

1. `test_cross_module_index_of_fncall` — verifies
   `make_arr(())[0]` resolves cross-module and returns 1.
2. `test_cross_module_string_interpolation_of_fncall` —
   verifies `IO.print("\(make_str(()))!\n")` produces
   `"hello!\n"` cross-module.

Both use the existing `_resolved` / `_compile_mod` /
`_run_mod` cross-module harness.  Both fail against the
pre-fix code path; both pass post-fix.

## Release prep (v0.0.147)

Bundled per the ABSOLUTE RULE:

- `vera/__init__.py` + `pyproject.toml` + `uv.lock`: 0.0.146 → 0.0.147
- `CHANGELOG.md`: dated `[0.0.147] - 2026-05-12` heading + link refs
- `HISTORY.md` Stage 12 row: v0.0.147 + roll-up counter 146 → 147
- `docs/index.html`: version tag
- `README.md` / `TESTING.md` / `ROADMAP.md`: count refreshes
  (3,815 tests, 147 releases)

## Verification

- mypy clean (59 source files)
- pytest: 3,801 passed, 14 skipped (was 3,799 + 2 new regression tests)
- All 86 conformance programs pass
- All 34 examples pass `vera check` + `vera verify`
- e602 gate: 116 files clean, 5 allowlist matched, 0 stale
- version-sync, doc-counts: consistent

Refs #614 #602 #626 #628

Co-Authored-By: Claude <noreply@anthropic.invalid>