feat: predicted run duration from historical task timings

[nam-hle]

Split out from the #370 "Ideas" parking lot (Task estimation).

Idea

Estimate how long a requested task graph will take before / during a run, using timing data recorded from previous runs.

Motivation

Nadle already records per-task durations (the profiling summary, critical-path, and cache-miss hotspots in --summary prove timing data is captured). That data could feed a forward estimate so a user knows roughly what they are about to wait for.

Possible shape (to refine)

• Persist per-task duration history alongside the cache (e.g. a rolling average per task id).
• Before a run, sum the estimated durations across the resolved DAG (respecting parallelism / worker count) to produce a total estimate.
• Surface it: e.g. a line at run start, and/or in --dry-run.
• Tasks predicted to hit cache contribute ~0; only cache-miss tasks add to the estimate.

Open questions

• Where to store history (cache dir? separate file?) and how much to keep.
• How to handle first-ever runs (no history) — omit the estimate vs show "unknown".
• Interaction with --parallel / worker count for a realistic wall-clock estimate.

Notes

• Spec-first per CLAUDE.md if this changes observable behavior/output.
• Relates to existing profiling output (--summary).

🤖 Generated with Claude Code