Non-interactive cost estimation and the scale escalation ladder

[jaakla]

Split out of #20 (workstream c).

Problem

#20 asks that very large analyses "try to estimate final runtime and ask user for confirmation". Two things block that as originally stated:

1. The CLI cannot block on stdin. openmapstack run is invoked non-interactively by subprocess.run in openmapstack/cli.py and by the clean-rerun protocol in openmapstack/rerun.py, and by CI. An interactive confirmation prompt would hang the harness and every eval that shells out to it.
2. An a-priori runtime estimate has no basis. There is no cost model. The only credible estimate is sampled wall-clock x scale factor, which requires Add partial run instructions and parameters for scalability #20's sampled run to have already recorded its scale factor and per-step timings.

Proposed split of responsibility

• CLI: emits a machine-readable projection and exits. Something like openmapstack run project.yaml --estimate --json, reporting projected wall-clock, projected output size, the scale factor it extrapolated from, and which sampled run record it derived them from. Never prompts.
• Agent/user layer: owns the confirmation. The skill decides whether to ask before committing to a multi-hour job.

Escalation ladder

When a job projects as very large, offer intermediate rungs rather than a binary accept/reject — city (typically the first test run), county, small country, large country, continent, global. Each rung is an ordinary sampled run from #20 with a larger AOI, so the ladder is a policy over an existing mechanism, not new machinery.

Acceptance criteria

• --estimate is non-interactive, machine-readable, and honest about its provenance: it must say it extrapolated from a specific sampled run, and refuse to guess when no sampled run exists rather than inventing a number.
• The escalation ladder is documented as skill guidance with the confirmation happening agent-side.
• Estimation error is reported as a range, not a false-precision point value.
• Tests cover: estimate with a prior sampled run, estimate with no sampled run (refuses, does not fabricate), and that no code path reads stdin.

Depends on

#20 — the sampled run record carrying scale factor and step timings is the input this consumes.