feat(gqt): add per-step object-store measurement under DST

[azimafroozeh]

What & why

Add --measure to omnigraph-gqt so existing .gqt cases report object-store work per step under DST. This makes request costs, repeated reads and work after publication visible when comparing engine changes.

• Record requests, repeated reads, bytes and requests after publication. Separate setup, case steps and runner checks, including steps with zero requests. Group requests by object-store operation and engine phase.
• Report phase timing and parallelism under --model unit|s3-like|local. Request counts participate in deterministic replay checks; timing, bytes and request logs remain measurement-only.
• Print tables and save a TSV alongside the invocation report. Accept multiple case files or directories from any location, with --artifacts <dir> to choose the output directory.
• Compare counts with an optional --baseline <path>. --write-baseline replaces only the measured case/environment/seed rows. Failed cases do not update their baseline rows; count differences are reported without failing the case.

Backing issue / RFC

Related to the operation-scoped cost evidence in the self-contained server testing RFC. No accepted backing issue or RFC is recorded for this change.

Checklist

• Change is focused on GQT measurement and reporting.
• Tests cover request accounting, phase timing, zero-request steps and baseline scoping.
• GQT README and developer testing docs updated.
• Author review against docs/dev/invariants.md.

Local verification

Recorded on the final implementation tree on 2026-09-28; not rerun for this description:

• cargo test -p omnigraph-gqt --locked: 174 unit tests, 197 corpus cases and 25 other runner tests passed.
• cargo clippy -p omnigraph-gqt --all-targets -- -D warnings: passed.
• cargo fmt --all --check, scripts/check-docs.py and scripts/check-agents-md.sh: passed.

git diff --check upstream/main...HEAD also passed while preparing this description.

Notes for reviewers

Measurement uses the existing object-store and decision seams in DST workers. Direct-engine environments do not record measurements. Latency and parallelism figures are model estimates, not wall-clock benchmarks. The engine and case format are unchanged; no benchmark cases or baseline data are included.