fix: make anti-slop advisory instead of auto-closing legitimate PRs (#741)

The `anti-slop` workflow was closing legitimate team-member PRs within
two minutes of opening. Root cause is a self-inflicted honeypot: the
repo's `pull_request_template.md` embeds an HTML comment instructing any
LLM/AI model to prepend the word "PINEAPPLE" to the PR description, and
`anti-slop.yml` lists that same word in `blocked-terms`. Any AI-assisted
PR from a team member follows the template instruction verbatim and then
trips the blocked-terms check — a guaranteed close. PR #739 (an
8-file, 496-line BQ finops fix with a passing consensus code review and
all other CI green) was auto-closed this way.

There is no `TEAM_MEMBERS` carve-out in `anti-slop.yml` the way
`pr-standards.yml` has one, so even listed members are subject to the
auto-close. The action's `authorAssociation` check only recognizes
GitHub's built-in OWNER/MEMBER/COLLABORATOR, not the repo's local
`.github/TEAM_MEMBERS` file.

Change: flip `close-pr: true` -> `close-pr: false`. All other anti-slop
behavior is preserved — failure labels still land (`needs-review:blocked`),
the failure message still comments on the PR, and the action still runs
on every open/edit/synchronize. Maintainers can close manually when the
signal is real. This keeps the anti-spam signal without the false-positive
close.

Two follow-ups are noted in the in-file comment and issue #740 but left
out of scope here:

1. Whether to keep `PINEAPPLE` in `blocked-terms` at all given that the
   template actively induces its inclusion.
2. Whether to add a `TEAM_MEMBERS` carve-out to anti-slop mirroring
   `pr-standards.yml`'s pattern.

Closes #740

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: anandgupta42

.github/workflows/anti-slop.yml

@@ -77,7 +77,13 @@ jobs:
             master
 
           # --- Actions on failure ---
-          close-pr: true
+          # Advisory, not blocking. Too many legitimate team PRs were auto-closed
+          # because the PR template instructs LLMs to prepend "PINEAPPLE"
+          # (honeypot) — any AI-assisted PR from a team member then trips the
+          # blocked-terms check and gets closed 2 minutes after opening. The
+          # labels and failure message still fire, so maintainers retain the
+          # signal and can close manually when warranted. See issue #740.
+          close-pr: false
           lock-pr: false
           failure-add-pr-labels: "needs-review:blocked"
           failure-pr-message: |