[zzylol]

Part 9 of the telemetry-functions stack: the three CI benchmark tiers from the design.

Tier	Workflow	When	What
Smoke	ci.yml	every PR and push to main	fmt, clippy, tests; then acceptance.py on real collector containers, with all 18 stages sharing the runner's 4 CPUs (a correctness check, not a capacity measurement); acceptance.json uploaded
Nightly	asap-sketch-benchmark.yml	daily	full rate sweep on vars.BENCH_RUNNER (falls back to ubuntu-24.04 with shared infrastructure); a new step compares against the previous night's summary.csv with benchmarks/compare-summaries.py and fails on a throughput drop beyond regression_margin (default 10%)
On demand	benchmark-on-demand.yml	a /benchmark PR comment by an owner, member or collaborator	reduced sweep on main, then on the PR head; posts the comparison table as a PR comment (the benchmark job runs PR code with read-only permissions; a separate job comments)

Also fixes the nightly's verify step, which still expected the 3-scenario row counts (189/63/441) after the top-k scenarios made it 5 (315/105/735).

To enable dedicated-CPU runs: register a self-hosted runner with Docker and gh, and set the repository variable BENCH_RUNNER to its label.

Validated with actionlint 1.7.7 (no findings) and yaml.safe_load. compare-summaries.py was tested on synthetic CSVs (−15% fails at a 10% margin; −5%, +4% and −1.2% pass). This PR's own CI run is the first real run of ci.yml.

Fixes found by this PR's first real CI runs

The shared 4-CPU runner exposed three problems that dedicated cores hid. All three were reproduced on a 56-core x86 server with every stage on CPUs 0–3 and fixed; the full acceptance then passed 25/25 under that contention.

1. Merge grace too short in scaled runs. A source's sketch sometimes reached window_merge after the scaled 2 s grace and was dropped as late. Acceptance now keeps the deployed 5 s grace, and the check lists any window whose counts differ.
2. The idle timer closed tumbling windows too early. Under contention it closed a window while its last batches were still queued, so sketch and exact windows differed by a batch (for example 388,736 vs 401,568 points). Data crossing a boundary already closes a window on time; the timer now closes only idle streams' windows, 5 s after their end.
3. Loki dropped one of two identical events. Loki keeps one entry per stream, timestamp and line. Sketch and exact novelty events had identical text (identical counts), so whichever arrived second was lost; on the runner that was the sketch's. Event lines now start with [sketch] or [exact].