…hemerals

The Vec-returning accessors (`Solver::resolved_packages` /
`resolved_ephemerals`, `ResolvePhase::solved_variants` /
`solved_ephemerals`) force callers that only need to iterate to pay
for an intermediate `Vec` plus, for ephemerals, an owning clone of
each `Requirement`. Add borrowing-iterator siblings so future
consumers can stream without allocating.

New API, additive — every existing Vec method keeps its signature:

  Solver::resolved_packages_iter()  -> Option<impl Iterator<Item = Rc<PackageVariant>>>
  Solver::resolved_ephemerals_iter() -> Option<impl Iterator<Item = &Requirement>>

  ResolvePhase::iter_solved_variants()   -> impl Iterator<Item = Rc<PackageVariant>>
  ResolvePhase::iter_solved_ephemerals() -> impl Iterator<Item = &Requirement>

`iter_solved_variants` still has to clone `self.scopes` internally
because `get_solved_variant` takes `&mut self` (it triggers a
deferred sort on the variant slice); the saving vs the Vec form is
the trailing `.collect()`. `iter_solved_ephemerals` is the bigger
win — pure borrow, zero allocation, no per-element clone.

Refactor the existing Vec methods to delegate to the iter forms so
there's one implementation of the filter logic.

## pyrer wired through

`crates/rer-python/src/lib.rs` switches its `SolveResult` build to
use `resolved_packages_iter` / `resolved_ephemerals_iter`. Two
intermediate `Vec`s skipped per solved result; for ephemerals every
entry is now read by reference instead of cloned then stringified.

## Tests

- `test_iter_resolved_packages_matches_vec_form` and
  `test_iter_resolved_ephemerals_matches_vec_form` in
  `solver::tests` — confirm iter and Vec forms agree on the same
  input and that the iter form returns `None` on a failed solve.
- Existing 5 Python tests for `resolved_ephemerals` still pass
  (the FFI surface is unchanged; just the implementation under it).

## Verification

- `cargo test` (Rust): 41/41 unit tests pass (was 39 + 2 new).
- `cargo test … --ignored` (188-case differential): 188/188 still
  match rez 1:1 in 17.68 s.
- `pytest tests/` (all Python): 80/80.

## Perf (188-case benchmark, same machine as README reference)

| | Total | Mean | Median |
|---|---:|---:|---:|
| README reference (post-#73)          | 11.35 s | 60 ms | 30 ms |
| This branch, pre iter-forms (run 1)  | 11.19 s | 60 ms | 28 ms |
| This branch, pre iter-forms (run 2)  | 11.27 s | 60 ms | 33 ms |
| This branch, post iter-forms (run 1) | 10.90 s | 58 ms | 28 ms |
| This branch, post iter-forms (run 2) | 11.16 s | 59 ms | 30 ms |

Within run-to-run noise; if anything a slight improvement from the
avoided allocations. No regression.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…hemerals

The Vec-returning accessors (`Solver::resolved_packages` /
`resolved_ephemerals`, `ResolvePhase::solved_variants` /
`solved_ephemerals`) force callers that only need to iterate to pay
for an intermediate `Vec` plus, for ephemerals, an owning clone of
each `Requirement`. Add borrowing-iterator siblings so future
consumers can stream without allocating.

New API, additive — every existing Vec method keeps its signature:

  Solver::resolved_packages_iter()  -> Option<impl Iterator<Item = Rc<PackageVariant>>>
  Solver::resolved_ephemerals_iter() -> Option<impl Iterator<Item = &Requirement>>

  ResolvePhase::iter_solved_variants()   -> impl Iterator<Item = Rc<PackageVariant>>
  ResolvePhase::iter_solved_ephemerals() -> impl Iterator<Item = &Requirement>

`iter_solved_variants` still has to clone `self.scopes` internally
because `get_solved_variant` takes `&mut self` (it triggers a
deferred sort on the variant slice); the saving vs the Vec form is
the trailing `.collect()`. `iter_solved_ephemerals` is the bigger
win — pure borrow, zero allocation, no per-element clone.

Refactor the existing Vec methods to delegate to the iter forms so
there's one implementation of the filter logic.

## pyrer wired through

`crates/rer-python/src/lib.rs` switches its `SolveResult` build to
use `resolved_packages_iter` / `resolved_ephemerals_iter`. Two
intermediate `Vec`s skipped per solved result; for ephemerals every
entry is now read by reference instead of cloned then stringified.

## Tests

- `test_iter_resolved_packages_matches_vec_form` and
  `test_iter_resolved_ephemerals_matches_vec_form` in
  `solver::tests` — confirm iter and Vec forms agree on the same
  input and that the iter form returns `None` on a failed solve.
- Existing 5 Python tests for `resolved_ephemerals` still pass
  (the FFI surface is unchanged; just the implementation under it).

## Verification

- `cargo test` (Rust): 41/41 unit tests pass (was 39 + 2 new).
- `cargo test … --ignored` (188-case differential): 188/188 still
  match rez 1:1 in 17.68 s.
- `pytest tests/` (all Python): 80/80.

## Perf (188-case benchmark, same machine as README reference)

| | Total | Mean | Median |
|---|---:|---:|---:|
| README reference (post-#73)          | 11.35 s | 60 ms | 30 ms |
| This branch, pre iter-forms (run 1)  | 11.19 s | 60 ms | 28 ms |
| This branch, pre iter-forms (run 2)  | 11.27 s | 60 ms | 33 ms |
| This branch, post iter-forms (run 1) | 10.90 s | 58 ms | 28 ms |
| This branch, post iter-forms (run 2) | 11.16 s | 59 ms | 30 ms |

Within run-to-run noise; if anything a slight improvement from the
avoided allocations. No regression.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>