…uests

Before, `aliasResolveHandler` matched raw absolute-path requests
(`hasRequestString === true`) only via `nameWithSlash`, which
unconditionally appends a forward slash. On Windows this meant a
raw request with the native backslash separator (e.g.
`C:\\abs\\foo\\baz`) failed `startsWith("C:\\abs\\foo/")` and the
alias was silently skipped. The `!hasRequestString` branch (file
hook) already used the normalized `absolutePath` and matched
correctly.

Mirror the `absolutePath` check into the `hasRequestString` branch
as an additive fallback — posix matches keep firing via the
existing `nameWithSlash` short-circuit, and windows backslash
subpaths now also match via the normalized form.

Tests

Adds a new `describe("absolute path aliasing (posix + windows
shapes)")` block that drives `aliasResolveHandler` directly
through a minimal `{ join, doResolve }` stub (the full-resolver
tests above use `memfs`, which is posix-only and cannot store
backslash paths). Coverage:

  posix:
    - raw absolute subpath request (raw-resolve hook)
    - joined absolute path (file hook)
    - exact equality match
    - negative: different prefix ("/abs/foo" vs "/abs/food")
  windows:
    - joined absolute path with native backslashes (file hook)
    - exact equality for raw windows absolute request
    - raw absolute subpath with native backslashes
      (the regression this commit fixes)
    - raw absolute subpath with forward slashes
    - negative: different prefix
    - negative: different drive letter

Confirmed the new windows-backslash-subpath test fails on the
pre-fix code and passes after; full suite is 1007/1007,
`npx tsc` and `npx eslint` clean.

https://claude.ai/code/session_013YaMjJMBeJYiu5WVfwEv2G

Replaces the hand-rolled `{ join, doResolve }` stub with a real
`new Resolver(...)` wired to the `AliasPlugin` under test. A
capture tap on the target hook records the aliased request and
ends the resolve, so the filesystem is still untouched and both
path dialects can run on the same machine — but every code path
now goes through the same `Resolver.doResolve` / `ensureHook` /
`resolveStep` plumbing the real pipeline uses.

Also factors the common alias-options setup into a
`createAliasRunner(options)` helper, one per `describe` group, so
each test only spells out the request shape it is exercising.

Re-verified by fault-injecting the pre-fix matcher into
`aliasResolveHandler` — only the windows-backslash-subpath test
fails, confirming the real-Resolver path still exercises the
regression. Full suite 1007/1007, tsc and eslint clean.

https://claude.ai/code/session_013YaMjJMBeJYiu5WVfwEv2G

Replaces the Resolver+capture-tap variant with a plain
`resolver.resolve(...)` call against the real `test/fixtures/`
tree, using `CachedInputFileSystem` (no filesystem stub) and
constructing every path through the Node `path` module.

Because `path.join` / `path.resolve` emit OS-native separators,
the same source covers both dialects on CI:

  Linux   →  `/absolute/imaginary-foo/index`  (posix)
  Windows →  `C:\...\imaginary-foo\index`     (backslash)

That means Windows CI exercises the raw-resolve-hook backslash
subpath case the previous commit fixed, without requiring any
platform-specific test code or a bespoke sync FS.

Cases covered:

  - raw absolute subpath request (raw-resolve hook).
  - request equals the alias name (exact equality).
  - relative request whose joined absolute path is aliased
    at the file hook after JoinRequestPlugin.
  - negative: a different absolute prefix sharing the same
    head must not accidentally fire (`imaginary-foo` vs
    `imaginary-food`).

The `imaginary-foo` directory is intentionally absent from
`test/fixtures/`, so every passing case must have gone through
the alias transform — if the matcher regresses the resolve fails
with ENOENT instead of silently landing on the wrong file.

Full suite 1001/1001; `npx tsc` and `npx eslint` clean.

https://claude.ai/code/session_013YaMjJMBeJYiu5WVfwEv2G