Two perf fixes from investigations/pacquet-macos-perf.md that both
collapse redundant passes over the tarball bytes.

1. Stream the HTTP body through the gzip decoder (doc fix #4).
   Replace `response.bytes().await` + `decompress_gzip` (sync
   zune-inflate) + separate `Integrity::check(&buffer)` with a
   streaming pipeline: `bytes_stream()` teed through
   `ssri::IntegrityChecker` into `async_compression::tokio::write::
   GzipDecoder`. This removes one full pass over the compressed body
   (no separate `.check()`), never holds compressed + decompressed
   buffers at the same time, and lets decompression overlap with
   network awaits. Integrity is still verified end-to-end before any
   CAFS blob is written, so a mismatch aborts the install with the
   same `TarballError::Checksum` behavior as before.

2. Interleave SHA-512 with the per-entry read (doc fix #2). Add
   `StoreDir::write_cas_file_prehashed`; the tar extraction loop
   reads + hashes 64 KiB chunks in one pass instead of
   `read_to_end` followed by a separate `Sha512::digest(buffer)`
   inside `write_cas_file`. The buffer still has to exist (CAFS
   filename is hash-derived), but two passes over the decompressed
   entry bytes collapse into one. Mirrors pnpm's `parseTarball` +
   `addBufferToCafs` in `store/cafs/src/addFilesFromTarball.ts`.

Why we diverge from pnpm on (1): upstream
`fetching/tarball-fetcher/src/remoteTarballFetcher.ts` +
`worker/src/start.ts` + `store/cafs/src/addFilesFromTarball.ts`
buffers the whole tarball into a Node `SharedArrayBuffer` and runs
`crypto.hash` + `zlib.gunzipSync` sequentially. Observable behavior
here is identical — same integrity errors on mismatch, same
decompressed bytes, same CAFS layout, same `index.db` rows — so per
the "Internal performance divergence is allowed" clause added in the
preceding commit, the streaming topology is a pacquet-specific
implementation-detail optimization rather than a shape port.

Dep changes: swap `zune-inflate` for `async-compression` (gzip +
tokio features) + `futures-util`; enable reqwest's `stream` feature;
add `sha2` to the tarball crate.

Tests: existing `packages_under_orgs_should_work` (real-network
end-to-end) and `should_throw_error_on_checksum_mismatch` both pass
unchanged, proving the streaming pipeline produces the same
decompressed bytes and fails integrity checks the same way.

Two perf fixes from investigations/pacquet-macos-perf.md that both
collapse redundant passes over the tarball bytes.

1. Stream the HTTP body through the gzip decoder (doc fix #4).
   Replace `response.bytes().await` + `decompress_gzip` (sync
   zune-inflate) + separate `Integrity::check(&buffer)` with a
   streaming pipeline: `bytes_stream()` teed through
   `ssri::IntegrityChecker` into `async_compression::tokio::write::
   GzipDecoder`. This removes one full pass over the compressed body
   (no separate `.check()`), never holds compressed + decompressed
   buffers at the same time, and lets decompression overlap with
   network awaits. Integrity is still verified end-to-end before any
   CAFS blob is written, so a mismatch aborts the install with the
   same `TarballError::Checksum` behavior as before.

2. Interleave SHA-512 with the per-entry read (doc fix #2). Add
   `StoreDir::write_cas_file_prehashed`; the tar extraction loop
   reads + hashes 64 KiB chunks in one pass instead of
   `read_to_end` followed by a separate `Sha512::digest(buffer)`
   inside `write_cas_file`. The buffer still has to exist (CAFS
   filename is hash-derived), but two passes over the decompressed
   entry bytes collapse into one. Mirrors pnpm's `parseTarball` +
   `addBufferToCafs` in `store/cafs/src/addFilesFromTarball.ts`.

Why we diverge from pnpm on (1): upstream
`fetching/tarball-fetcher/src/remoteTarballFetcher.ts` +
`worker/src/start.ts` + `store/cafs/src/addFilesFromTarball.ts`
buffers the whole tarball into a Node `SharedArrayBuffer` and runs
`crypto.hash` + `zlib.gunzipSync` sequentially. Observable behavior
here is identical — same integrity errors on mismatch, same
decompressed bytes, same CAFS layout, same `index.db` rows — so per
the "Internal performance divergence is allowed" clause added in the
preceding commit, the streaming topology is a pacquet-specific
implementation-detail optimization rather than a shape port.

Dep changes: swap `zune-inflate` for `async-compression` (gzip +
tokio features) + `futures-util`; enable reqwest's `stream` feature;
add `sha2` to the tarball crate.

Tests: existing `packages_under_orgs_should_work` (real-network
end-to-end) and `should_throw_error_on_checksum_mismatch` both pass
unchanged, proving the streaming pipeline produces the same
decompressed bytes and fails integrity checks the same way.

- Move `configure_rayon_pool` to after `CliArgs::parse()`. clap's
  `--help` / `--version` paths exit inside `parse`, so they never
  reach the rayon pool init now — `pacquet --help` no longer
  spawns worker threads it'll throw away. Other subcommand paths
  still pay the (small) init cost; we can move it into the install
  path specifically as a follow-up if non-install commands are
  ever added that don't need rayon.
- Drop the local `CasPathsArc` / `WarmEntry` aliases in
  `create_virtual_store::run`. They duplicated the value type of
  `pacquet_tarball::PrefetchedCasPaths`; if that map's value type
  changes, the local aliases would silently drift. Letting `warm`'s
  element type be inferred from its `push` calls keeps the binding
  to the upstream type.
- Add three unit tests for `prefetch_cas_paths`:
    * `prefetch_cas_paths_returns_hits_for_live_index_rows` — happy
      path: index row written, CAFS blob present, verify on; result
      contains the expected file map.
    * `prefetch_cas_paths_omits_failed_integrity_entries` — index
      row whose digest matches no on-disk file; verify on; row is
      dropped from the result rather than surfacing a half-populated
      map.
    * `prefetch_cas_paths_skips_filesystem_checks_when_verify_disabled`
      — same shape as the failing-integrity test but with
      `verify_store_integrity = false`; the row surfaces (no fs
      checks), matching pnpm's `verify-store-integrity: false`
      semantics.

#4 (cmd-shim — chmod target swallowed all errors): the post-write
\`ensure_executable_bits\` call used \`let _ = ...\` which dropped
\`PermissionDenied\`, \`EROFS\`, AppArmor deny and other real failures
alongside the intentionally-ignored \`NotFound\`. Match \`NotFound\`
explicitly and propagate everything else as \`LinkBinsError::Chmod\`,
mirroring pnpm's \`fixBin\` ENOENT guard. Adds two DI-driven
regression tests:

- \`link_bins_propagates_target_chmod_error_via_di\` — \`PermissionDenied\`
  must surface as \`Chmod\`.
- \`link_bins_swallows_target_chmod_not_found_via_di\` — \`NotFound\`
  must remain swallowed.

#5 (package-manager — peer-resolved slot name parsing): \`find_slot_
own_package_dir\` used \`slot_name.rfind('@')\` to strip the version
tail. That works for \`parent@1.0.0\` and \`@scope+parent@1.0.0\` but
breaks for peer-resolved slots: \`to_virtual_store_name\` produces
shapes like \`ts-node@10.9.1_@types+node@18.7.19_typescript@5.1.6\`
where the last \`@\` is inside a peer's version, not at the
package-name boundary. \`rfind\` then split inside the peer spec, the
slot's own package wasn't found, and the slot was silently skipped —
bins of children of every peer-resolved slot were never linked.

Parse from the left instead, skipping a single leading \`@\` that
belongs to a scoped package. The package-name half can never contain
\`@\` after \`to_virtual_store_name\`'s scope normalization (\`/\` →
\`+\`), so the first non-leading \`@\` is the right boundary. Adds
\`link_virtual_store_bins_handles_peer_resolved_slot_name\` using the
exact \`ts-node@10.9.1_@types+node@18.7.19_typescript@5.1.6\` shape
verified by \`pacquet_lockfile::pkg_name_ver_peer::tests::to_virtual_
store_name\`.

Both fixes verified by re-breaking the relevant code path and
confirming the regression tests fail, then reverting.

Drive-by from #9 (style nit on the chain split in
\`SymlinkDirectDependencies::run\`): the chain was broken in the
earlier \`feat(cmd-shim): port directories.bin discovery and Windows
shim formats\` commit (dfdd33a) because \`direct_deps\` is consumed
twice — once by \`symlink_direct_deps_into_node_modules\`, once by
\`link_direct_dep_bins\`'s \`dep_names\` derivation. That can't sit
inside a single chain. Documenting it here in lieu of amending the
earlier commit.

Test count: cmd-shim 70 → 72, package-manager (lib) 38 → 39.

The Windows CI runner caught two test-only Windows regressions
introduced by the shell-selection change earlier in this branch:

- `extend_path::tests::pnpm_virtual_store_layout_yields_three_bins_deepest_first`
  hard-coded `/proj/...` expected values. `path::absolute("/proj")`
  on Windows resolves against the current drive (`C:\proj`), making
  those expectations racy with the runner. Split into a Unix-gated
  test with literal expectations and a new platform-neutral test
  that asserts the structural invariants (count + deepest-first
  ordering) without anchoring to any absolute root.
- `lifecycle::tests::lifecycle_emits_script_stdio_and_exit_in_order`
  and the new `child_sees_stamped_npm_package_and_no_leaked_npm_config`
  both use POSIX-only script syntax (`;`, `1>&2`, `printf`, `$VAR`).
  Before this branch the spawn used a hard-coded `sh -c` so Git
  Bash on the Windows runner satisfied them; once the shell
  selector started picking `cmd /d /s /c` on Windows (per item #4
  / parity with pnpm) those tests can no longer run there. Gate
  them on `cfg(unix)`. The cmd-path shell selection itself is
  covered by `crate::shell` unit tests; env stamping by
  `crate::make_env` units.

No production-code change.

…ecycle scripts (#397) (#418)

## Summary

Closes three of the critical items in #397 by porting the corresponding
behaviors from `@pnpm/npm-lifecycled2d8e790` and `pnpm/pnpm@b4f8f47ac2`:

- **#1 — Lifecycle env vars.** New `make_env` module ports `makeEnv` and
  the surrounding env block in `lifecycle()` ([npm-lifecycle/index.js:73-104](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/index.js#L73-L104), [:354-414](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/index.js#L354-L414)) plus the pnpm wrapper's `extraEnv` additions ([runLifecycleHook.ts:119-124](https://github.com/pnpm/pnpm/blob/b4f8f47ac2/exec/lifecycle/src/runLifecycleHook.ts#L119-L124)). Lifecycle scripts now see `npm_lifecycle_event`, `npm_lifecycle_script`, `npm_node_execpath`/`NODE`, `npm_package_json`, `npm_execpath`, `npm_package_*` (`name`, `version`, and recursive `config`/`engines`/`bin`), `npm_config_node_gyp`, `npm_config_user_agent`, `INIT_CWD`, `PNPM_SCRIPT_SRC_DIR`, and `TMPDIR` (when `unsafe_perm` is false). The spawn now strips inherited env (`env_clear()`) so leftover `npm_*` keys from a wrapping invocation cannot leak through.
- **#2 — PATH ancestor walk.** New `extend_path` module ports `extendPath` ([npm-lifecycle/lib/extendPath.js:5-27](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/lib/extendPath.js#L5-L27)) plus the tri-state `scriptsPrependNodePath` gating ([:29-61](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/lib/extendPath.js#L29-L61)). For a dep at `<root>/node_modules/.pnpm/<slot>/node_modules/<pkg>`, `PATH` now contains the dep's own `.bin`, the `.pnpm` slot's `.bin`, the project root's `.bin`, the bundled `node-gyp-bin` (when supplied), `extra_bin_paths`, and finally the inherited PATH.
- **#4 — Shell selection.** New `shell` module ports the [shell-selection block](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/index.js#L241-L252) and the [pnpm-side `.bat`/`.cmd` guard](https://github.com/pnpm/pnpm/blob/b4f8f47ac2/exec/lifecycle/src/runLifecycleHook.ts#L63-L71). `cmd /d /s /c` on Windows, custom `scriptShell` on either platform, otherwise `sh -c`. The Windows-batch-file `scriptShell` case surfaces as `ERR_PNPM_INVALID_SCRIPT_SHELL_WINDOWS` (matching upstream's error code).

`RunPostinstallHooks` grows seven new fields to surface these knobs; `BuildModules` passes safe defaults (`None` / `true` / `Never`) for all of them — full config plumbing for `user-agent`, `unsafe-perm`, `scripts-prepend-node-path`, `node-gyp` bundling, and `script-shell` are tracked as separate items in #397 (#14, #15) or follow-ups.

### Explicit non-goals in this PR

Three caveats called out in #397 that are deliberately deferred:

- `windowsVerbatimArguments` (Rust equivalent: `CommandExt::raw_arg`) is signalled by `SelectedShell.windows_verbatim_args` but not yet applied to the spawned `Command`.
- `@yarnpkg/shell` / `shellEmulator: true` has no clean Rust port; pacquet ignores the flag for now.
- `unsafe_perm` uid/gid drop (#14) — `BuildModules` passes `true`, which keeps current behavior (no TMPDIR creation, no privilege drop).

### Test parity

Per the project guide ("port the relevant pnpm tests too whenever they translate"), this branch ports:

- `test('makeEnv')` from [npm-lifecycle/test/index.js:97-124](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/test/index.js#L97-L124).
- The `extendPath` ordering test from [npm-lifecycle/test/extendPath.test.js:5-8](https://github.com/pnpm/npm-lifecycle/blob/d2d8e790/test/extendPath.test.js#L5-L8).
- `runLifecycleHook() does not set npm_config env vars` from [pnpm/exec/lifecycle/test/index.ts:65-77](https://github.com/pnpm/pnpm/blob/b4f8f47ac2/exec/lifecycle/test/index.ts#L65-L77), adapted to a file-dump model so we don't need the IPC fixture.
- `onlyOnWindows('pnpm shows error if script-shell is .cmd')` from [pnpm/exec/commands/test/index.ts:509-542](https://github.com/pnpm/pnpm/blob/b4f8f47ac2/exec/commands/test/index.ts#L509-L542) and the custom-shell case from [:478-508](https://github.com/pnpm/pnpm/blob/b4f8f47ac2/exec/commands/test/index.ts#L478-L508).

Plus fills the gaps where upstream coverage is thin: tri-state `scriptsPrependNodePath`, two-level pnpm virtual-store PATH walks, `extra_bin_paths` ordering, TMPDIR gating on `unsafe_perm`, and `extra_env` precedence vs `npm_lifecycle_script`.