perf(share): gather provider sessions in parallel

[od0]

Layer 2 of the picker-performance stack (on #156, under #159 → #154).

The unified session picker (path share, bare path resume in #154) enumerated the seven providers sequentially, so the slowest scan — a big codex or claude history — stacked on top of all the others. gather_artifacts now fans the provider scans out to scoped threads: wall time is max-of-providers instead of the sum.

• Claude scans inline on the calling thread — ClaudeConvo caches its chain index in a RefCell, so it cannot cross threads. Codex dominates the wall clock in practice, so this costs nothing.
• Ordering preserved: results concatenate in the old sequential provider order before the stable sort, so ranking tie-breaks are byte-identical. A panicking collector degrades to that provider missing from the picker, matching existing warn-and-continue handling.

Measured (with #156 beneath): gather drops from ~8.5 s to ~5 s debug on a 421-rollout tree; release lands ~2.5 s cold. #159 above this takes warm opens to milliseconds.

path-cli 0.16.1 → 0.16.2. Version slots: #138/#145 also claim numbers in this range — whoever lands second renumbers.

[github-actions]

🔍 Preview deployed: https://c072734d.toolpath.pages.dev

[od0]

Reasoning from Claude Code on the changes in this file:

The why: the listing cache in cmd_share needs the exact same stat-level enumeration the sync engine uses (that's the whole invariant — one enumeration path, so cache stamps and sync stamps can never
disagree). The tuple structs are private to sources.rs, and rather than making the structs themselves pub(crate) — exposing their internals — the cache work added three small constructor functions
returning impl ArtifactSource, only for the three providers the cache consumes (claude/codex/opencode; that's why gemini/cursor/pi/copilot lines are untouched).

The line-53 edits you're looking at are then source_for delegating through those new functions instead of invoking the structs directly. Strictly optional — same module, it could still call
ClaudeSource(...) — but it keeps exactly one construction site per provider, so a future change to how a source is built can't silently diverge between the sync engine and the cache. Three lines of
diff noise bought for that guarantee; I think it's the right trade, but if you'd rather source_for stay untouched, reverting those three lines is behavior-neutral.

[akesling]

I think a match over a loop of ArtifactType would probably be more immediately readable. While this makes adding a new line much simpler, understanding the actual logic requires a new indirection and reading a macro definition.

This would also make the spawn and join shapes match, decreasing mental overhead for bookending.

e.g.

let mut types = /* all the types here */
if let Some(exclusive_harness) = harness_filter {
    types = Vec::from([exclusive_harness]);
}
for t in types {
    handles.push(s.spawn(match t {
        ArtifactType::Gemini => move || {
            collect_gemini(mgr, canonical_cwd, canonical_project.as_deref()) }
       ....
    }))
}

Remember that collect_gemini, etc. are just helpers for use here and you can make them return an easier to use return type, etc.