test: make cold-refresh performance snapshots robust (Fixes #509) (#510)

## Summary

Make cold-refresh quality snapshots statistically robust by measuring
ten independent cache-cold refreshes and ten paired cache-warm
refreshes. Cold P50 now catches systematic regressions without allowing
one host-cold event to dominate the gate; warm P50/P95 remain
independently enforced.

## Changes

- use a unique cache directory for each cold/warm pair and fresh PET
processes for both members
- keep existing `full_refresh` and time-to-first fields as warm
compatibility metrics
- add schema-v2 cold refresh/time-to-first distributions and
independently cold phase/locator diagnostics
- enforce cold P50 with calibrated platform dual budgets once both
snapshots are v2
- use explicit fail-closed absolute ceilings while the exact base has
legacy schema-v1 metrics
- require all cold diagnostics, stable inventories, and privacy-safe
timeout counts
- document workload semantics, migration behavior, and measured budgets

## Validation

- `cargo test -p pet --test e2e_performance` (6 passed)
- local release cold/warm benchmark completed with 10 samples per
workload
- `python -m unittest discover -s scripts/tests -p 'test_*.py' -v` (31
passed)
- `cargo test --all`
- mandatory format/clippy pre-commit checks
- unchanged-head all-platform runs `31509651276` and `31510433777`
passed
- both real schema-v2 runs pass against exact schema-v1 main baselines

Fixes #509

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: karthiknadig