[Perf] nvm use: skip re-resolving an exact, installed version

[hridhinvarkeyjoseph]

Fixes #3930.

Problem

On every new shell, nvm.sh calls nvm_auto use. With no node on PATH yet, nvm_auto resolves the default alias to an exact installed version (e.g. v24.21.0) and then calls:

nvm use --silent v24.21.0

nvm use then resolves that already-exact version from scratch, twice:

1. VERSION="$(nvm_match_version "${PROVIDED_VERSION}")" → nvm_version
2. nvm_ensure_version_installed "${VERSION}" ... → nvm_version again, then nvm_is_version_installed

Each nvm_version call does alias resolution plus nvm_ls, forking several subshells. The existing code already flagged this spot:

# This nvm_ensure_version_installed call can be a performance bottleneck
# on shell startup. Perhaps we can optimize it away or make it faster.

Profiling (master, clean env -i environment, avg of 10 calls)

call	time
nvm use --silent v24.21.0	168ms
nvm_ensure_version_installed v24.21.0	57ms
nvm_version v24.21.0	48ms

About 105ms of the 168ms is spent re-deriving a version that is already exact. Changing my default alias from 24 to v24.21.0 made no difference (~290ms either way), so exact versions go through the full resolution path too.

Change

Adds nvm_is_installed_exact_version, a cheap check that returns success only when the argument:

• is exactly vX.Y.Z (digits and dots only, so v12, v12.0, v12.0.0-rc.1, node, lts/*, iojs-v1.0.0, system, etc. do not qualify),
• is not shadowed by an alias of the same name (aliases take precedence in nvm_version, so this preserves that), and
• is installed, using the existing nvm_is_version_installed (executable bin/node, or node.exe on Windows; respects the old directory layout via nvm_version_path).

When it passes, nvm use sets VERSION directly and skips nvm_match_version and nvm_ensure_version_installed. Everything after that is unchanged: PATH/MANPATH, NVM_BIN/NVM_INC, hash -r, the current symlink, nvm_die_on_prefix, --save, and output.

All other inputs (partial versions, aliases, lts/*, --lts, system, iojs, .nvmrc, and versions that aren't installed) go through the existing code path, so their behavior and error messages are identical.

Results

Sourcing nvm.sh in a clean environment (env -i), macOS arm64, zsh, one installed Node version, default alias = 24 (avg of 4 warm runs):

	startup
master	~330ms
this PR	~235ms

That's roughly 95ms (~29%) saved on every new shell. Users with more installed versions or slower fork should see a larger absolute saving, since the skipped code calls nvm_ls.

Tests

• New: test/fast/Unit tests/nvm_is_installed_exact_version covers: no argument; exact version not installed; non-executable node binary; installed exact version; a range of non-exact inputs (12.0.0, 12, v12, v12.0, v12.0.0-rc.1, node, stable, lts/*, iojs-v1.0.0, system, default); and an alias named v12.0.0 taking precedence over the fast path.
• Existing: nvm_is_version_installed, nvm_ensure_version_installed, and all Running 'nvm use …' fast tests pass in sh, bash, and zsh.
• Running 'nvm use system' should work as expected fails on my machine, but fails identically on unmodified master (I have no system node installed), so it's unrelated to this change.
• I couldn't verify ksh locally: macOS ships ksh93, which has no local, so the existing tests fail there as well. CI should cover it.

Manual check:

$ nvm use v24.21.0   # fast path
Now using node v24.21.0 (npm v11.19.0)
$ nvm use 24         # existing path
Now using node v24.21.0 (npm v11.19.0)
$ nvm use v99.0.0    # not installed: existing path, same error, exit 3
N/A: version "v99.0.0" is not yet installed.

[hridhinvarkeyjoseph]

Pushed a fix for the bash-only failure: the new helper needed to be added to nvm unload's function list. All fast bash jobs now pass.

The one remaining failure (WSL nvm install (Ubuntu-24.04, 16)) timed out during wsl --install, before any nvm code ran, and the matching (Ubuntu-24.04, 16, script) job passed, so it looks like a runner flake. Could it be re-run when convenient?

[hridhinvarkeyjoseph]

Folded in the change from #3945 (skip nvm_version in nvm_resolve_local_alias for an exact, installed version), since both are the same fast path on the startup route — one PR to review instead of two.

[ljharb]

i tweaked the second commit to use the helper the first commit introduced.