gcec: drop the unfounded "CI green twice" bar from the fallback-coverage task

task.md's step 5 required CI to go green twice, justified as "it touches
test:live". That trigger doesn't match the rule it borrowed from and has no
rationale behind it:

- RULES.md scopes the twice bar to e2e/heavy-browser changes — test:e2e over
  dev/requirements/heavy/**, real Chrome, the genuinely flaky category the rule
  exists for.
- test:live is live.test.js, extractor-support.test.js and
  fallback-coverage.test.js, which run offline against committed cached HTML.
  live.test.js's own header: "deterministic and runnable anywhere (no network)."

So the routine was imposing a second full CI run on a deterministic suite. Drop
the sentence and let RULES.md stay the single source for when the twice bar
applies; the skill link right beside it already routes to the merge flow.

Found while merging #755, where I cited this line as a blocker it wasn't.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XyQTydqacfHB2Rkq9aKvXs

Author: claude

.claudinite/local/packs/gcec/tasks/fallback-extractor-improvements/task.md

@@ -104,8 +104,8 @@ bash .claudinite/local/packs/gcec/tasks/fallback-extractor-improvements/postcond
 Branch `claude/fallback-coverage/<date>`, commit the change plus the regenerated
 GENERATED artifacts, and push. Then open a **PR for review** — it never merges
 itself; from there it goes through the usual flow (the gcec pack's
-[merge-and-ci skill](../../skills/merge-and-ci/SKILL.md)), and CI must
-go green **twice** (it touches `test:live`). PR body: the hypothesis, why it's generic (which unseen sites
+[merge-and-ci skill](../../skills/merge-and-ci/SKILL.md)). PR body: the
+hypothesis, why it's generic (which unseen sites
 it helps), how the jsdom trap was ruled out, the before→after numbers, and the
 covering test.