Profile Poetry identification and avoid whole-inventory cloning with an alias index

Start with crates/pet-poetry/src/lib.rs at find_with_cache and try_from, then use the scale fixtures from #533 and ownership guidance from #536. Measure identification, allocation, operation counts, and timings across representative inventory sizes before choosing whether indexing is justified. Done means evidence-backed behavior tests preserve alias, symlink, workspace, refresh, and invalidation semantics, or document a measured deferral.

Tracking plan: #528
Priority: P3, measurement-gated optimization. Evidence: cloning/linear scan confirmed; end-user latency impact not yet profiled.

Problem

Poetry's cached discovery lookup clones the complete LocatorResult, including all environments, and try_from then scans that result and its symlinks for each candidate. At larger inventories this can produce repeated whole-inventory allocation and candidate-by-environment work. Cold concurrent callers can also repeat cache population work.

Sources: find_with_cache, try_from.

Scope

First measure identification and allocation/operation counts over representative Poetry inventory sizes using deterministic fixtures. If material, share immutable discovery results and index executables/aliases for direct lookup. Reuse existing per-key single-flight machinery where appropriate instead of adding another cache framework.

Preserve Poetry-versus-generic-venv precedence, prefix/project/manager fidelity, platform path semantics, and the full/workspace/kind-filtered refresh-state merge contract. Do not introduce broad cross-refresh cache persistence without an invalidation design.

Acceptance criteria

• The issue records before/after operation or allocation counts and timings across multiple inventory sizes; any optimization claim is tied to evidence.
• If indexing is justified, repeated identification no longer clones/scans the entire inventory for each candidate.
• Alias/case/symlink identity, workspace association, manager details, refresh merging, and configuration invalidation have behavior tests.
• Concurrent cache misses avoid unnecessary duplicate work without deadlocking or permanently caching failures.
• Small-inventory latency and memory do not materially regress under #531/#533.
• If the measured benefit is negligible, close with measurements and a documented deferral instead of landing complexity solely to remove clones.

Dependencies

Depends on #533 for scale fixtures and #536 for stable cache/snapshot ownership. This is deliberately after the reproduced process/transport failures; do not treat it as the first performance fix.