Two cross-file dedups plus stale-allow removal:

- **`push_normalized_component`** existed twice — `path.rs` and
`cache/cached_path.rs` — identical except the cached-path copy trims the
trailing `\0` that uvwasi appends to directory entries on wasm
([nodejs/uvwasi#262](nodejs/uvwasi#262)). This
keeps one `pub` wasm-aware version in `path.rs`. Note this means
`path.rs` callers (tsconfig path joins) now also get the `\0`-trim on
wasm; on native targets the generated code is unchanged, and the helper
stays `#[inline]` so the hot `normalize_with` path keeps its codegen.
- **`manual_tsconfig()`**: the "manually configured tsconfig,
deliberately skipping `Auto` discovery" match block was duplicated
verbatim in `lib.rs` (`resolve`) and `dts_resolver.rs`
(`dts_resolve_tsconfig_paths`). One helper in `tsconfig_resolver.rs` now
owns it, and `find_tsconfig_manual` drops its `pub(crate)`.
- **Stale allows**: `clippy::cognitive_complexity` on `extend_tsconfig`
and `clippy::too_many_lines` on `dts_try_extensions`' neighbor no longer
suppress anything (verified by removing them and running the CI clippy
command); `extend_tsconfig` keeps `too_many_lines` (139/100).

Part of a second cleanup pass; follows #1264–#1268.

…#1270)

The two "resolve from symlinks" benches did, per timed iteration, 10,000
eager `format!("./file{i}")` allocations plus an `assert!(.. .is_ok())`
branch — unlike every sibling bench in the file, which iterates
precomputed data and discards results with `_ =`. The asserts also
re-checked an invariant already validated once before the benchmark
group runs.

This precomputes the specifiers once (mirroring the `find tsconfig`
bench's hoisting pattern) and drops the in-loop asserts, so the benches
measure symlink resolution rather than resolution + allocation noise.
The validation pass also now constructs one resolver instead of 10,000.

**Heads-up:** this will step-change the CodSpeed series for
`resolver_memory/resolve from symlinks` and `resolver_real/resolve from
symlinks` downward — that's the removed allocation/assert overhead, not
a resolver perf change.

Part of a second cleanup pass; follows #1264–#1268.

Small leftovers found while sweeping the test suite and the JS side:

- `src/tests/resolve.rs`: the "file in module with query and fragment"
table row appeared twice, byte-identical — the upstream enhanced-resolve
`resolve.test.js` has the case exactly once, so the second row was a
copy-paste in the port. The loop just ran the same assertion twice.
- `src/tests/alias.rs`: `#[allow(clippy::too_many_lines)]` no longer
suppresses anything (verified by removing it and running the CI clippy
command). The same allow in `imports_field.rs` is still load-bearing
(1191/100) and stays.
- `vite.config.ts`: `"napi/browser.js"` was listed twice in
`fmt.ignorePatterns`.
- `napi/test.mjs`: removed a stale comment about a `.ts` extension next
to code that adds `.mjs` and resolves an already-suffixed specifier.
- `napi/webcontainer-fallback.js`: removed an inert lint directive (`//
eslint-disable-next-line: no-console`) — the project lints with oxlint,
the syntax is invalid even for ESLint, and `no-console` isn't enabled.

Part of a second cleanup pass; follows #1264–#1268.