eval: report partial Alibaba dashboard measurements

[zzylol]

Why

Preserve the real Alibaba dashboard measurements collected before the user requested stopping the experiment.

What

Add the Rust comparison harness, data preparation and evaluation design, plus 15 completed service runs out of 54 planned runs. Include raw plans/measurements, dataset provenance, a partial report and performance plots. Execution is stopped; this is not the completed experiment.

How

Prepare all 240 official three-minute archives (12 hours, 2,646,269,187 retained observations). Calibrate on a 10M-event reservoir from hours 0–6; replay hours 6–12 without sampling after one hour of warmup. Compare six baselines on count and Top-3 dashboards with 1m/10m/60m windows refreshed every minute.

Before this PR

The dashboard comparison branch had no executable Alibaba observation-contract experiment or these real-data measurements.

After this PR

Reviewers can inspect the partial report, plots and raw per-query samples, including planning, update, merge, readout, payload memory and accuracy violations. This PR is stacked on #602.

Trial 0 total CPU seconds:

Baseline	Update	Merge	Readout	Accuracy violations / 2160
Exact-pane	66.95	10.35	4.07	0
Exact-scan	7.90	0	2270.60	0
Analytical sizing	1076.88	10.39	9.68	1102
AutoSketch CPU extension	3287.68	10.72	1.15	378
ERP-NoSharing	3289.36	10.69	1.15	378
ERP shared-or-local → Exact fallback	67.03	10.23	3.93	0

Verification

• All 15 completed runs pass artifact checks for event counts, endpoints, plan identity, timers, violations and shared-merge accounting; completed exact references have zero accuracy violations.
• Seven Rust correctness tests passed before execution; preparation tests are separate from real performance evidence.
• Four reporting-validation tests pass, including a regression for insignificant JSON floating-point round-trip differences; integer configuration fields remain exact.
• No performance rerun was started after the stop request. The interrupted AutoSketch third run is excluded.
• Product screenshots: not applicable; measured plots are included instead.

Limitations

Service repetitions are incomplete/unbalanced. Edge and latency catalogs exist, but their held-out runs were not started. The displayed table is consistently trial 0, not a three-trial aggregate. ERP uses a custom calibration catalog, not synthetic nearest-shape matching; the shared ERP result is exact fallback, not a successful approximate-sketch result. AutoSketch is a CPU adaptation with extensions. Analytical sizing is not full cost-model optimization. Timers measure operator regions, not deployed backend/network latency; memory is logical payload, not isolated RSS. Original datasets and oracle caches are not vendored into Git.