Accepts the after_step: ADR-0008 amendment

[johnnyt]

Flips the 2026-09-08 after_step: amendment in
docs/adr/0008-durable-subchart-child-runs.md from proposed to accepted,
now that sp-c48 is on main at e3209bd (PR 91). Docs-only: one
Markdown file under docs/adr/, no lib/, no test/, no fragment.

What changed

• The amendment's Status line flips in place: proposed -> accepted (2026-09-08, sp-nhl; ... flipped once sp-c48 landed). Those two lines
  are the only removed lines in git diff origin/main -- docs/adr/.
• A dated Note at the foot of the file records the merge and what was
  re-read against main at 3fd45f7 before the flip. Amend-by-addition:
  nothing above it is reworded.

Verified before the flip (against main at 3fd45f7 = e3209bd + the 0.11.0 prep)

Every clause of the amendment, its worked example and its "what this
section does not decide" hold as written. By anchor:

• Clause 1, the whole effect list. report_step/3 is reached from
  defp persist_tail(store, run_id, machine_state, effects, executor, write, reporter) in runs.ex with that function's own effects
  parameter; Enum.split_with/2 never rebinds it, so the reporter gets
  the pre-split list. DriverTest's "hands over the whole effect list,
  lifecycle effects included" asserts a {:done, _} reaches the callback
  and not the executor.
• Clause 1, no public return widened. The seam landed as a
  package-internal step_reporter: option in StatifierPersistence.Runs'
  t:opt/0 union, documented there as this package's own and never a
  host's. No public Runs function's return changed; the arity changes
  are all private (persist_tail/6 -> /7, stepped/6 -> /7).
• Clause 2, two entry points. The driver calls exactly three Runs
  functions - Runs.create/4 from create/3, Runs.step/5 from defp step(driver, run_id, opts, event, ref), and Runs.cascade_cancel/3,
  which steps nothing. The two step-taking ones are the two the clause
  names, and the driver drains the reporter's buffer and fires the host's
  callback from those two sites after the entry point returns.
• Clause 3's ordering and clause 4's propagating raise hold, the
  latter with a test ("a raise inside the callback propagates to the
  caller").
• Every function-head anchor the amendment cites still resolves in
  driver.ex and runs.ex. The one cite whose spelling moved is clause
  1's persist_tail head, which gained the reporter argument; the Note
  says so.

Recorded in the Note rather than edited into the text

• The per-call override landed as Keyword.get(opts, :after_step, driver.after_step), not the Keyword.put_new/3 spelling the closing
  section named - the same rule ("the caller's option outranks the
  driver's field") written from the other end, since the default lives on
  the struct.
• A discarded delivery fires nothing, refused on both sides of the
  seam, with a test whose recorded sabotage needed both halves broken.
• Clause 3's lock statement holds by construction and no test asserts
  it: serialized/5 closes before the entry point returns and the
  callback fires after that return. Said plainly rather than papered over.
• The amendment's own "this section carries no lib/ change and no test,
  and waits for one" is falsified by the flip; per the record-flip rule it
  is met in the Note, not reworded.
• The seam section's "decision 5 below says why" reads as clause 5 -
  the amendment's own fifth clause. This record's decision 5 is the
  cascade, and the section's other self-references say "clause N". Left as
  written, corrected in the Note (residue the pass-2 reviewer of PR 89
  left).

Gate

Not run, and none applies: the diff touches one Markdown file under
docs/adr/ and no Elixir code - the authority table's "a change touching
no Elixir code has no gate to run and may commit on review of the diff
alone". No changelog fragment: changelog.d/README.md puts documentation
and ADRs on its "do not write a fragment" side.

Closes sp-nhl.

[johnnyt]

Direction review pass 1: UNQUALIFIED, 0 findings. Merged --rebase at d72bbf6.