`load_alias_by_options` re-ran `compile_alias` (allocating and cloning
every `AliasValue`) on **every failed resolution**, while the primary
`alias` list is compiled once at construction. This compiles `fallback`
once into a `ResolverImpl.fallback` field, uses `load_alias` directly at
the fallback call site, and deletes `load_alias_by_options` (its only
caller). `compile_alias` is pure over the post-`sanitize` options, so
behavior is identical; with an empty fallback list the new field is an
empty box + 32 zero bytes.

Also tidies `clone_with_options` while touching it: the `(a && !b) ||
(!a && b)` yarn_pnp cache check becomes a plain `==` comparison, and the
stale `#[allow(clippy::unused_self)]` is removed (both cfg branches use
`self`).

Split out of #1263.

)

`Cache::get_tsconfig` deep-cloned every parsed `TsConfig` before
branching on `root`, but the clone only matters in the root branch,
where `TsConfig::build(mut self)` consumes the value after the raw
version is cached. On the non-root path — taken once per `extends` entry
— the original local was dropped unused, so the clone
(`CompilerOptions`, three `Option<Vec<PathBuf>>`s, references, paths)
was pure waste.

The raw cache still stores the `should_build = false` version and the
built cache still stores the built version; returned `Arc`s share the
cached instances exactly as before.

Split out of #1263.

…de (#1266)

- Factor the byte-identical 14-line `typings`/`types`/`main`
entry-selection block, which appeared twice in
`dts_resolve_as_directory` (once in the `typesVersions` branch, once in
the types/typings/main branch), into a `dts_package_entry` helper, and
collapse the resulting nested `if let` into a let-chain.
- Remove `impl BitOr for Extensions`: the `|` operator is never invoked
on this private type — every union goes through `.union()` (`BitAnd` is
used and stays).
- Drop the unused `_cached_path` parameter from
`dts_resolve_tsconfig_paths` (private method, single caller).

Split out of #1263.

A batch of tiny, independent cleanups:

- Derive `Clone` for `Resolution` — the manual impl was field-for-field
identical to what the derive produces (`Debug`/`PartialEq` stay manual,
they're genuinely custom).
- Merge the two byte-identical `#[cold]` empty-error constructors in
`Specifier::parse` into one.
- `require_hash`: `.ok_or_else(..)` instead of `.map_or_else(|| Err(..),
Ok)`.
- Remove a pointless `Some(owned).as_deref().unwrap()` dance in the
legacy `jest-runner-../..` fallback path.
- Drop a redundant `.as_ref()` on `file_name()` in `Cache::value`
(`&OsStr == &str` already works via `OsStr: PartialEq<str>`).
- Remove the napi `EnforceExtension::{is_auto,is_enabled,is_disabled}`
helpers — zero callers anywhere in the repo and not `#[napi]`-exported,
so they never reached JS.
- Remove a no-op `let _ = meta;` in the `metadata` test.
- Fix the stale doc comment on `TsConfig::root` / `root()": it claimed
the field drives final template-variable substitution, but that's keyed
off `should_build` now; `root` just records whether the config is the
caller's (vs loaded through `extends`).

Split out of #1263.

)

The "return the last array entry's error" branch in
`package_target_resolve` was dead code: `i` comes from `enumerate()`, so
`i == targets.len()` could never be true. Errors from array entries were
always swallowed like `undefined`, so an invalid target in the last
fallback entry surfaced as a generic `PackagePathNotExported` instead of
the specific error — despite the comment right below claiming "Return or
throw the last fallback resolution null return or error".

This fixes the off-by-one (`i + 1 == targets.len()`) so the last entry's
error now propagates, matching the documented intent and Node's
`lastException` behavior for arrays.

One deliberate exemption: `ResolveError::NotFound` still falls through.
It only arises from re-resolving a bare `imports` target (e.g. `\"#a\":
[{\"import\": [\"#b/import.js\"]}]`) through the module system, and
enhanced-resolve treats an unresolvable candidate there as a soft
failure, not a hard error — the ported "Direct and conditional mapping
#4" test pins that behavior. Node would actually throw immediately on
any non-`ERR_INVALID_PACKAGE_TARGET` error inside the array loop, which
is stricter than enhanced-resolve; the exemption keeps us compatible
with the latter.

Adds two imports-field tests: an invalid target in the **last** array
entry now errors (previously `Ok(None)`), and an invalid target in a
**non-last** entry still falls through to the next entry.

Split out of #1263.