fix(discovery): prune stale entries by pid liveness, not shared-port reachability

[runyourempire]

The bug

victauri check / invoke / coverage return 401 Unauthorized against a running, auth-enabled app on any machine that has run more than one Victauri app.

Root cause: victauri-test's discovery (find_live_servers) inferred a server was alive purely from TCP reachability of its advertised port. But every Victauri app registers on the same default port (7373). So when a previous app crashes/exits, its stale <pid>/ discovery dir still advertises 7373 — and a different live app now holding 7373 makes that port probe succeed. The dead entry is never pruned, and its stale token is added as a "live server". Two entries then match port 7373, so unique_token_for_port returns None (ambiguous) → no token sent → 401.

This is why the MCP-bridge path (victauri-cli::bridge, which already gates on the pid) worked while check/invoke/coverage didn't — the exact asymmetry.

The fix

Gate liveness on the owning pid, not just port reachability: a dead owner is stale even when the shared port answers, so the entry is pruned and token selection stays unambiguous.

• find_live_servers → testable find_live_servers_in(base, is_alive, is_reachable) + cross-platform is_process_alive (tasklist / kill -0), mirroring the audited helper in victauri-cli::bridge.
• diagnose_discovery gates Live/Stale on pid liveness too (accurate classification on a shared port).
• Also resolves victauri doctor reporting "server not running" in this case (it was the same failed discover()).

Tests

New regression test: two discovery entries on the shared port (one dead, one live) → only the live token is selected and the dead dir is pruned — the exact 401 scenario, now green. Plus a mid-rebuild case (live owner, unreachable port → pruned).

cargo test -p victauri-test 115/0, cargo clippy clean (deny-level workspace lints), rustfmt clean, and the repo pre-push gate passed.

Notes / not in scope

• Reproduced live while introspecting a sibling app: 5 stale %TEMP%\victauri\<pid> entries all on 7373 caused the 401; pruning dead-pid dirs fixed it. Discovered via the stale installed CLI too (0.5.6) — unrelated; upgrading is the separate fix.
• Doesn't change the shared-default-port design itself (7373 for all apps) — it just stops stale entries on that port from poisoning token selection.

🤖 Generated with Claude Code

[runyourempire]

Note on the two red checks — pre-existing, not from this PR.

This change touches a single file (crates/victauri-test/src/discovery.rs) with no Cargo.toml/Cargo.lock changes, so it can neither cause nor fix dependency-advisory checks:

• Security Audit — already failing on main (92e854d): RUSTSEC-2026-0185 (quinn-proto) + unmaintained atk (RUSTSEC-2024-0413), transitive deps.
• Dependency Checks — RUSTSEC-2026-0190 (unsoundness in anyhow::Error::downcast_mut), a 2026 advisory published after this branch's base ran, so it newly trips cargo-deny's live advisory DB. Fixed by bumping anyhow — a separate dependency-hygiene PR.

The checks that validate this change are Check & Lint (fmt/clippy/cargo test) and E2E Integration; the regression test (dead_pid_sharing_a_live_port_is_pruned_so_selection_stays_unambiguous) covers the exact 401 scenario.