ci: establish dedicated nightly perf validation (#2085)

[justinjoy]

Summary

Establish the dedicated Linux nightly performance owner before changing the
hosted PR validation contract tracked in #2085 and blocking #2077.

• Run Linux performance work only on [self-hosted, Linux, X64, perf],
  scheduled or manually dispatched on canonical main, with a shared queue.
• Use the runner's local semantic-reasoning:ubuntu26 Docker image without
  pulling it; execute the inspected image ID and record its identity.
• Require nine actual CRDT/CSPA trials, gold correctness, CoV eligibility,
  host readiness, and absence of CPU throttling. Preserve existing targets.
• Report acquisition completeness separately from the performance verdict;
  skips, missing samples, and target failures do not become a green result.
• Prebuild measurement binaries and prevent test-time rebuilding. Preserve
  broader TRACE suite, portfolio, DOOP, and Windows coverage by default.
• Provide an explicitly partial strict_smoke dispatch for validating the
  CRDT/CSPA acquisition path without running the entire nightly workload.

Evidence and sequencing

Hosted calibration runs 37288534027 and 37307436532 produced no raw timing on
9 of 12 VM observations. Measured identical PR code passed or failed the old
CRDT absolute threshold depending on CPU model; all measured CSPA samples
missed the old absolute threshold. This does not establish a source regression
in #2077.

This PR leaves the hosted required check unchanged and does not close #2085.
After merge, manually dispatch Perf Nightly on main with strict_smoke=true
and require complete, eligible acquisition. An existing target miss remains a
failed performance verdict. Then the next PR can establish hosted correctness
checks with strict performance checks owned by nightly, and #2077 can be rebased
and its required CI rerun.

Validation

Focused Docker, authentic host telemetry, measurement-log, Meson target
selection, governor, DOOP, Windows, TSan, and release RSS contracts are checked
locally. Python checks use a uv virtual environment under HOME.

The actual dedicated runner/local-image acquisition must be verified after
merge: this environment has no Docker daemon. The installed actionlint does
not yet recognize GitHub's concurrency.queue; its other diagnostics are
checked separately.

CI failure follow-up

The first GCC build failed the Python encoding gate and Bash construct ratchet.
The original commit was amended to specify UTF-8 in all 11 reported operations,
including the inherited hosted_perf_calibration.py operation on main, and to
write the telemetry fixture using individual printf calls that the static
shell checker can resolve. Neither gate was weakened.

Local revalidation: encoding scan (115 files, zero findings), seven encoding
self-tests, Bash construct self-test and real setup-only Meson introspection
(78-file closure, zero findings), eight Docker contracts, release RSS contract,
and real capture/check telemetry fixtures including rejection cases all pass.
Full build/ABI validation is being rerun by GitHub CI on the amended commit.