feat(telemetry): measure the seconds of each row #4058

Problem

No metric gives the seconds of one row. The job seconds counter has the labels repo and lane. It is counted once for a lane round (server/service.py, _execute_batch). So the 5 minute rule of AGENTS.md cannot be checked for one row. A slow row hides in the lane total.

Expected

Add one histogram devops_cli_roadmap_row_seconds with the labels repo, row and result.

• _execute_due_rows (run.py) times each row around row.runner(...) and observes the seconds.
• result uses the same values as devops_cli_service_jobs_total (_count_job).
• The label is row, not job. The collector sets the label job from the resource and loses a data point attribute of that name (tool_metrics.py).
• A row that is not due, or that the ledger suspends, adds no sample.

Series count: repo x row x result. The status, triage and model lanes hold at most 11 rows after the issue #4052. Each row has the results of _count_job. The bound is 11 rows x 3 results = 33 series for one repo, times the histogram buckets of the instrument.

Document the series in docs/TELEMETRY.md.

This issue changes no label of the model usage metrics. It keeps the label scheme of #3069 (repo, item_class, result) and of #3666 (repo, lane). It uses repo, row and result in the same meaning, so a dashboard (#3655) can join them.

No live model run (AGENTS.md rule 13).

Acceptance criteria

• uv run pytest tests/telemetry/test_roadmap_row_seconds.py -k row_seconds - a round emits one sample for each row that ran, with its row and result.
• uv run pytest tests/telemetry/test_roadmap_row_seconds.py -k not_run - a row that is not due, and a row that the ledger suspends, add no sample.
• uv run pytest tests/telemetry/test_roadmap_row_seconds.py -k error_result - a row that raises adds a sample with the error result.
• grep -q 'devops_cli_roadmap_row_seconds' docs/TELEMETRY.md - the documentation lists the series (it prints nothing today).

Every command above fails on the unchanged code. The pull request cites the fail-to-pass run (devops ci fail-to-pass --base <ref>).

Feasibility

I read these files at origin/main: server/service.py (_execute_batch), telemetry/instruments.py (emit, InstrumentKind.HISTOGRAM), roadmap/tool_metrics.py (the pattern and the row label), run.py (_execute_due_rows, _count_job) and docs/TELEMETRY.md. I did not run the code. The change makes no GitHub request.

Existing tool considered

telemetry/instruments.py (emit), tool_metrics.py as the pattern, and the OpenTelemetry histogram that exists. No new library.

Cost

Build: about 30k tokens and 1.5 h. Keep: about 2k tokens per quarter. Fix: about 8k tokens and 0.5 h.

Related

Part of #3027. #3069, #3650, #3655, #3666 (see the comment on #3069). Order: after the issue #4052 (both change run.py).