perf(alias): parallelize prewarm git/config/user-config reads

[max-sixty]

Three commits that compound. Together they cut the alias-dispatch warm critical path from ~38ms to ~12ms by overlapping the three independent pre-dispatch reads that previously ran in series.

What was slow

For wt <alias>, the pre-dispatch path had three blocking reads in sequence:

1. git rev-parse --git-common-dir ... (prewarm, ~7ms warm / ~10ms cold)
2. git config --list -z (all_config, called via project_identifier_warm, ~10ms warm)
3. UserConfig::load_with_warnings (TOML parse of wt.toml, ~4ms cold)

Standalone measurement showed git config --list -z is dominated by git's process startup (~4ms of ~5ms total — git --version alone is ~4.6ms), and the same applies to the rev-parse fork. Adding a dependency like gix-config would only buy back the subprocess overhead. Running these reads concurrently buys it without a new dep.

How it works

Repository::prewarm_at runs three sibling threads inside a std::thread::scope:

• prewarm_rev_parse — populates GIT_COMMON_DIR_CACHE etc. (unchanged from before)
• prewarm_git_config — runs git config --list -z and stashes the parsed IndexMap in a new GIT_CONFIG_PRELOAD static
• prewarm_user_config — calls UserConfig::load_with_warnings and stashes the result in WORKTRUNK_USER_CONFIG_PRELOAD

Repository::at consumes both preloads into the per-Repository RepoCache (all_config and user_config OnceCells), so the first access from a Repository is a memory hit. On-demand fallbacks remain in place for tests and any caller that bypasses prewarm.

LoadedConfigs::load was a sequential pair of cache reads after this — both fields became thin cache hits. The third commit inlines it at all 8 callsites and folds the surviving invariants (warning ordering, why user/project stay distinct) into Repository::project_config's rustdoc. Net −62 lines.

Naming

The git/worktrunk-config split is now visible at a glance: GIT_* for git-side data (GIT_CONFIG_PRELOAD, GIT_COMMON_DIR_CACHE, GIT_DIRS), WORKTRUNK_* for worktrunk-side (WORKTRUNK_USER_CONFIG_PRELOAD).

Numbers

cargo bench --bench alias:

• dispatch/warm/1: ~13.8ms → ~11.5ms (-17%)
• dispatch/warm/100: ~17.4ms → ~12.7ms (-27%)
• dispatch/cold/1: ~14.5ms → ~11.6ms (-20%)

wt-perf timeline -- stub shows the three sibling spans inside the prewarm scope starting within ~22µs of each other.

Behaviour change worth flagging

User-config deprecation warnings (and the "↳ To see details, run wt config show" hint) now emit on every wt invocation in a deprecated-config repo, not just commands that previously called repo.user_config(). Process-wide dedup (WARNED_DEPRECATED_PATHS for warnings, DEPRECATION_HINT_EMITTED for the hint) keeps it to once per process. Behaviour is now uniform across commands; before, wt config show rendered deprecations into stdout but didn't emit them on stderr, while wt list did.

This exposed and removed a duplication in wt config update: format_update_preview and the --print path were emitting warnings via format_deprecation_warnings directly, which now races prewarm and prints duplicates. Drop the redundant emission — prewarm is the single canonical source. Project-config warnings still emit via check_project_config's separate check_and_migrate call.

Minor wart, not blocking: the hint says "to apply updates, run wt config update" even when the user is running that exact command. Fixable later by reading argv before prewarm or moving the hint into command handlers.

Behaviour nuance

prewarm_git_config runs git config --list -z from discovery_path instead of git_common_dir (the rev-parse thread is racing in parallel and git_common_dir isn't known yet). For the default config (extensions.worktreeConfig off — the common case) the output is byte-identical: linked worktrees and the common dir share one config file. With extensions.worktreeConfig enabled, the new code reads per-worktree values that the old code masked; that's arguably the correct behaviour when a user opts into worktree-scoped config, and worktrunk doesn't itself enable the extension.

Testing

Standard pre-merge gate (wt hook pre-merge --yes) passes — 3464 tests including config-show snapshot tests with the new stderr lines. 18 snapshot files updated to reflect the warning-uniformity change; net diff is one or two added stderr lines per test.