Checkpoint long work so a stop costs minutes, not the day

Situation

A task runs for hours. The session that holds it can be cut off by a context limit, a crash or a person closing the laptop, and everything the worker learned since the last durable record goes with it. A transcript is not a record; it is not resumable and it is not the tool's to keep.

What you run

• checkpoint: a short body on stdin becomes a capped record under the local half, with identity computed from git
• check: reports whether the checkpoint is fresh against the profile's interval, beside scope and blockers
• context: the next worker reads the newest checkpoint inside the assembled context, not a pasted log

Scenario

setup: active-task
given:
  - 'an active task scoped to lib, with work done inside that scope'
steps:
  - id: checkpoint
    run: ['checkpoint']
    stdin: checkpoint-body.md
    note: 'a capped progress record; its branch and head are computed from git, never written by hand'
    expect:
      exit: 0
      stdout_contains: ['\.ai/local/state/checkpoints/']
  - id: fresh
    run: ['check']
    note: 'the checkpoint is fresh against the profile interval and the task is consistent with git'
    expect:
      exit: 0
      stdout_contains: ['0 failing']
  - id: resume
    run: ['context']
    note: 'the newest checkpoint is part of what the next worker reads'
    expect:
      exit: 0
      stdout_contains: ['^## TASK', 'checkpoint']
then:
  - 'the checkpoint file carries the branch and head it was written at'
  - 'check reports the checkpoint age against the profile interval'
  - 'context names the task and its newest checkpoint'

Outcome

Progress exists as a file the tool can find, capped so that it stays a summary, and the profile decides how stale is too stale. A session that stops mid-task loses at most one interval.