Activity body: the definition collapses in parts, and its identity always arrives

[m2ux]

Summary

When a resumed worker asks for its activity again, almost everything in the response can arrive as a
short marker instead of the bytes: each bundled technique, the inherited rules block, each shared
block of a composed technique, each eagerly bundled resource. One part could not. The activity
definition was concatenated onto the response outside the delivery ledger, so no lookup happened and
no marker was possible — it was appended in full every time, to a worker still holding the previous
copy. Measured over eight activities of the main workflow, that was 70,957 of 184,684 characters,
38.4% of what a resume received.

Work item W10 of the delivery-cost epic #404. Stacked on #441, which this branch includes.

The identity is what makes the fix a split rather than a marker

A worker is required to confirm that the activity id the server returned matches the one it was
dispatched for, and to stop without executing a step if they disagree. That check reads the
definition. So the definition cannot collapse to a single marker the way a technique does.

It is keyed in parts instead. The identity — every scalar field of the definition — is always
delivered whole. The step list, the transitions, the outcome, and the artifact contract the server
synthesises from the steps' declared outputs are keyed separately, by content, under
activity:<field>:<hash>. That is the treatment a composed technique already gets, where the
invariant note and the item list are keyed apart so a shared preamble collapses even when the rest
differs.

Content-keying means no invalidation: an edited step list hashes differently, so it gets a different
key and delivers in full.

A field the splitter does not recognise stays with the identity, so a definition field the server has
never seen is delivered rather than silently dropped.

What it measures

The reference walk over thirteen gates of the main workflow, one identity per activity:

	Before	After
Re-request collapse	69.9%	83.5%
Characters re-requested	271,411	141,951

The epic body quotes 65.4% as the figure to beat. That reading predates the batched-dispatch merge,
which added the activity-worker role technique to every activity delivery; the same test on the
branch point reports 69.9%, and that is what the table above is read against.

Acceptance, item by item

• A resumed worker's re-request returns the identity in full and the remainder as markers where the
  bytes are unchanged. Asserted on the real corpus: id: present, steps: a marker, response
  shorter.
• The dispatched-activity confirmation still has an id to check on every delivery path, including a
  fully collapsed one. Asserted by parsing a definition whose every keyed section collapsed and
  reading id and version off it.
• A forced full delivery is unaffected. Asserted with bundle: 'full' after a first delivery.
• The reference walk records a higher collapse figure, with the body share falling. 83.5% against
  69.9%, above.

Scope of change

• src/utils/activity-body.ts (new) — the split and the per-field projection.
• src/tools/workflow-tools.ts — the delivery path reads the projection; the cost line reports how
  many definition fields collapsed and what they were worth.
• src/utils/delivery.ts — the new namespace joins the ledger's key inventory.
• tests/activity-body-delivery.test.ts (new, 9 tests) — the split loses nothing, an unrecognised
  field survives, a changed section delivers in full while its siblings collapse, and the two
  corpus-level assertions above.

No schema change, no definition change.

Non-goals

• The eager bundle's budget behaviour. A marker costs nothing against the budget, so freed headroom is
  spent sending resource bodies the first delivery could not afford — that is the budget working, not
  re-delivery, and it is why a large resume figure can sit beside a small body share.
• Re-warming a resumed worker's context, which is a consequence of gates waiting on a person.
• Confirming on production runs that re-delivered characters have fallen. That is W11, and it waits
  on around ten completed gate-crossing sessions; the walk test is a regression guard, not evidence
  about production.

Investigation detail

engineering/artifacts/planning/2026-08-02-delivery-cost-epic/issue-417-activity-body-delivery.md
— the per-activity body-share table, from 17% on an activity that binds many techniques to 95% on one
that binds few.