pr_review.py: a GitHub Actions billing-failure notice can be misread as UNRECOGNIZED_REVIEWER_OUTPUT #1071

What happened

Driving ptr727/Financial-Modeling#199, pr_review.py wait reported:

UNRECOGNIZED REVIEWER OUTPUT (1): ...
  body carrying no heading at all, which no measured review body does  (round 49064e1b)
status=UNRECOGNIZED_REVIEWER_OUTPUT ...

Direct GraphQL confirmed the flagged body:

copilot-pull-request-reviewer, submitted 2026-08-28T23:06:34Z: "The job was not started because
recent GitHub Actions payments have failed or your spending limit needs to be increased."

This isn't a code review at all - it's GitHub Actions' own billing-failure notice (the same one
blocking every workflow job on the repo at that moment, confirmed separately via gh run view's
own annotations), somehow posted as a PR review body attributed to the Copilot reviewer account
rather than surfacing only as a run annotation. The timestamp lines up exactly with the billing
outage window on this repo (recent payment failure / spending limit).

Why this matters

The script's own caution here is correct - it should not trust a digest it cannot fully parse - but
its message doesn't distinguish "an unrecognized review shape" from "not a review at all, an
infra-failure artifact." The two need different next steps: a genuinely new review body shape is a
parser gap to fix; a billing-failure notice masquerading as a review is closer to the class #1060
already tracks (checking budget/quota status before reporting a reviewer bot as broken), just
inverted - here a billing failure got misattributed as reviewer output instead of causing an
absent one.

Suggested fix

Recognize this specific shape (or the general "no code-review markers, matches a known
GitHub-infra-failure phrasing" pattern) and report it distinctly, e.g. status=REVIEWER_INFRA_FAILURE
with a pointer to check the repo's Billing & plans settings, rather than folding it into the generic
UNRECOGNIZED_REVIEWER_OUTPUT bucket that asks the agent to file an issue every time it recurs.

What I did instead

Manually verified via gh api graphql (querying reviews.nodes with commit.oid) that: (1) the
flagged body is the billing artifact above, dated exactly during the confirmed outage window, (2)
the most recent real review from the same reviewer (copilot-pull-request-reviewer, later commit)
reads "Approval recommended... no remaining correctness gaps identified", and (3) reviewThreads
shows 0 unresolved. Proceeded to merge on that basis, per the script's own "whether to merge anyway
is the maintainer's call" - this note is that call, made with the evidence in hand rather than by
overriding the caution unread.