feat: a reviewer's surplus insight becomes the next item, not a refusal (#171)

[thedancingdeveloper]

Two items were rejected having met every criterion their briefs stated, both
because the reviewer would additionally have done something else:

the other two bundle entry points have no explanation at all

the code does not demonstrate any explicit guard against future accidental
serialization of Snippet::text

Both are good observations. Neither was in the item. Rejecting on them
discarded the work and the observation — the item goes back to be rewritten
identically, and nothing records what was noticed. Both have since been filed
by hand as rdpapp#2 and rdpapp#3, which is the behaviour this makes
automatic.

What changes

When every stated criterion is met, the reviewer approves and lists what else
it would have done under Follow-ups. Those are kept as plan-shaped markdown
for a person to accept or discard, plus an event each — the stream is the
source of truth, and a coordinator should not need to know where a file is.

The safety property

Only on approval. A rejection has already said what is wrong and gets no
second channel. Allowing follow-ups there is exactly how "approve and defer"
grows into a way to wave a failed criterion through — the gate-weakening
AGENTS.md forbids, arriving by a friendlier route.

The prompt says it plainly:

This does not soften anything below. A criterion the task states and the
change does not meet is a rejection, and no follow-up substitutes for one.

and a test asserts a rejection produces nothing at all.

Two smaller boundaries, both tested

• "- none" is an answer, not an item. A backlog filled with the absence of
  findings is worse than an empty one.
• Prose after the list ends the list. Attributing a proposal the reviewer
  did not make is worse than missing one: a person triages it, finds nothing
  behind it, and trusts the next one less.

Why a plan document rather than queue rows

The same reason inception produces a PLAN.md: writing straight to the queue
forks the pipeline into a generated path and a hand-written one that diverge
forever. Nothing here is admitted to anything until a human moves it into a
plan.

Both executors share the parser and the recorder, so this cannot drift the way
the review rubric did (#167).

All four gates green.

Closes #171.

🤖 Generated with Claude Code