fix(review-loop): accept a verdict that closes a summary paragraph

[rlorenzo]

The bug

prompts/ and lib/ disagreed about where the verdict goes.

The prompt asked the reviewer to "state in the Summary: **Verdict: good to go**". The parser required a line of its own:

^[[:space:]]*([-+][[:space:]]*)?verdict[[:space:]]*:[[:space:]]*good to go[[:space:]]*[.!]?[[:space:]]*$

A reviewer that follows the prompt and ends its summary paragraph with the verdict produces a report that reads as unclean.

Observed on a real run. An antigravity review closed with:

… Finding counts: High 0, Medium 0, Low 0. Verdict: good to go.

test_review_clean never matched, so the loop could not break early. It ran all five cycles against an already-clean diff, spent 7m54s re-reviewing work no agent had touched, printed MAX ITERATIONS REACHED - some issues remain, and wrote a needs-review receipt for bin/review-gate to act on. Every one of the six passes had reported High 0 / Medium 0 / Low 0.

The fix

lib/lib-review-loop — the verdict may now open a line or close a sentence at the end of one. Both guards the anchors existed for survive by construction:

• Quoting prose — "the report says Verdict: good to go" — still fails, because says is not a sentence boundary.
• Hedged prose — "good to go, but …" — still fails, because the trailing anchor allows only an optional . or !.

The leading marker class also repeats now, so a ## Verdict heading counts; it did not before.

Prompts — code-review and code-review-followup now say where the verdict goes instead of leaving it to the reviewer. The loop drives five agent CLIs and none can be relied on to place it, so this is defence in depth — the parser is the fix. The code-review wording is generated, so the edit is in .claude/commands/code-review.md; tools/generate carried it to prompts/ and the four vendor SKILL.md copies.

Test plan

• ./test/run → 182/182 pass (was 181 + 1 stale-generation failure caught mid-work by generate --check).
• Four new regression tests in test/lib-review-loop.bats: the paragraph-closing verdict that failed, a ## heading verdict, and two pinning the guards against hedged and quoted verdicts.
• Replayed the real agent-code-review.md from the failing run through both regexes — old: no match (confirms root cause); new: clean.