test: measure client-observed refresh latency (Refs #531) (#542)

## Summary

Measure refresh latency at the JSONRPC client boundary instead of
treating server discovery duration as total request latency.

Refs #531; part of #528. Initial hosted calibration is now recorded in
`31551ea`; this draft remains open for CI and review before closing
#531.

- Record client round-trip, request-relative time to first environment,
process-start-relative time to first environment, and server discovery
duration separately, including cold distributions.
- Preserve exact-base schema-v2 semantic comparisons; never compare old
discovery-only samples as client RTT. Schema-v3 artifacts require all
new distributions, and v3-to-v3 comparisons activate empirically
calibrated client gates without changing existing
discovery/startup/coverage budgets.
- Exercise the real `PetClient.refresh`
serialization/write/buffered-read path using a deterministic channel
fixture: 80ms/50ms client RTT despite 2ms/1ms server duration, with TTFE
resetting to 40ms/7ms. Malformed environment notifications fail before
timestamps/inventory are recorded.

## Validation

Local Windows validation:

- `cargo test --offline --locked -p pet --test e2e_performance`: **8
passed**.
- `python -m unittest discover -s scripts\tests -p 'test_*.py' -q`: **55
passed**.
- `scripts\rust-precommit.ps1`: formatting and workspace
warnings-as-errors Clippy passed.
- `cargo clippy --offline --locked --workspace --all-targets
--all-features -- -D warnings`: passed.
- `git diff --check`: passed.
- Independent repository Reviewer follow-up confirmed
malformed-notification feedback resolved.

## Draft gates and compatibility

Calibration now uses three successful hosted baseline runs plus the
original PR measurement at identical source `9c1b003`, giving 40
cold/warm pairs per platform. Run links, stable inventories, observed
ranges, and new client-only budgets are documented in
`docs/QUALITY_SNAPSHOTS.md`; replay tests cover unchanged-source
variance and material regressions. Client absolute budgets retain at
least twice observed statistic ranges; existing
discovery/startup/coverage gates are unchanged. The v2-to-v3 transition
retains old gates and explicitly reports new client metrics as
diagnostics when the exact base lacks comparable data.

Environment notifications lack request IDs. The measured summary uses
one refresh per fresh process to avoid post-response/concurrent
notification ambiguity; no general concurrent TTFE attribution is
claimed. Existing buffered stdout reading and bounded stderr draining
are retained.

This PR is independent of #541's runtime glob fix and does not include
#529's EOF changes. Preserve both changes when integrating the
benchmark-client shutdown work later. No merge or auto-merge requested.

---------

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

Author: karthiknadig