Activity exits: the investigation record behind the work package

[m2ux]

Summary

The investigation record behind the activity-exits work package: the route counts the change was
measured against, the naming decisions it rests on, and the per-site tail measurements that picked
which exits become immediate.

One new file, no edits to anything existing.

Why it lands first

#501 pins .engineering at this commit. Until it is on the engineering branch, that pointer names
a commit reachable only from a feature branch — nothing fails at checkout or build, since the object
lives in this repo, but .gitmodules tracks engineering, so a git submodule update --remote
would quietly move off it. Merging this before #501 keeps the pointer honest.

Both PR bodies in the stack cite this record for their measured claims, so it is also where a
reviewer goes to check the numbers rather than take them on trust.

What it records

• Route counts per workflow before the move — 176 declarations across 117 activity files in 17
  workflows, broken out as 127 transitions (48 conditional), 15 decision branches and 34
  checkpoint-option routes, with a per-workflow table. And zero after.
• Why the count falls by 13 rather than staying flat — nine decision branches that named no
  destination and so answered nothing about where next, six checkpoint options that reached a
  destination an exit already reached, and two activities that gained a named ending where they had
  been falling off the end of their routes.
• The reconciliation with the issue's figures. Activity exits: the workflow stitches the graph, not the activities #496 cites 208; the corpus had drifted to 176 by
  the commit this branch starts from, and the file and workflow counts match exactly.
• The exit-naming table — the thirty routes whose names could not be derived from a predicate,
  each with the name chosen for it, plus the four derived names that read as the negation of a flag
  and the two branch ids that named the activity they routed to.
• The tail measured after every routed checkpoint option — how many steps follow it and how many
  of those are ungated. This is the evidence for which five aborts become immediate exits, and for
  leaving the two hand-gated review dispositions alone.

Non-goals

A record of what was measured, not a design document. The design is stated in the definitions and in
the PR bodies; this exists so the numbers in them can be checked.

🤖 Generated with Claude Code