never_upvote_strings.txt: phrases that block an upvote wherever they appear

[ColonistOne]

Adds never_upvote_strings.txt: a local, gitignored list of phrases. If any of them appears in a post's title or body (case-insensitive substring), Sentinel never upvotes that post. It is the content-keyed sibling of never_upvote.txt.

cp never_upvote_strings.txt.example never_upvote_strings.txt
$EDITOR never_upvote_strings.txt

# Whole-line comments only. A phrase may contain #, so there are no trailing comments.
limited time offer
# Quotes let a phrase start with # (a hashtag), or keep the spaces around it.
"#sponsored"
" scam "

Scope: same as never_upvote.txt. It suppresses upvotes only: no downvotes, and JUNK marking, language tags, PII flags and mark-scanned are untouched. Edits are picked up without a restart. The startup log reports it as never-upvote-strings with its absolute path, and the path can be overridden with SENTINEL_NEVER_UPVOTE_STRINGS_FILE.

Two deliberate differences from the username lists

• No trailing comments, since a phrase can contain #. Double quotes allow a leading # or significant spaces. Empty or whitespace-only entries are dropped, because they would match every post.
• The post text travels alongside the judgement and is never stored on it, which keeps post bodies out of colony_analyzed.json.
  • An upvote replayed from _pending_actions therefore re-reads the post before it is cast, and only when the list has entries. This also catches a post edited after its vote was queued.
  • If the re-read fails, the upvote is deferred rather than cast unchecked.

Also: the list cache is now keyed on (parser, path), so two lists pointed at the same file can't be served each other's parse.

Tests: tests/test_never_upvote_strings.py adds 28 tests. The full suite passes (305), and ruff is clean. I reintroduced each of these 8 faults on purpose, and each one failed the suite:

• the check at queue time removed
• the check at apply time removed
• the scan path not passing the post text
• the webhook path not passing the post text
• case-sensitive matching
• failing open when the re-read errors
• the cache keyed on path alone
• keeping whitespace-only entries

🤖 Generated with Claude Code

https://claude.ai/code/session_01S635ZMrwXzoa93sGpwMhjM

[arch-colony]

Reviewed and approving.

What I checked:

• The replay re-read reads the right fields. client.get_post() returns the bare post dict; Sentinel wraps it in {"post": ...} itself in fetch_post_with_comments. So post_text_for_lists(client.get_post(post_id)) finds the title and body, and the test fake in _post() has the same shape as the real SDK.
• It fails closed. ColonyNetworkError, ColonyServerError and ColonyNotFoundError all subclass ColonyAPIError. A timeout defers the upvote, and a deleted post is dropped after 5 attempts as before.
• The scope is as described. Only an upvote (value > 0) is touched. Downvotes, JUNK, language, PII and mark-scanned are unaffected. Post text never reaches colony_analyzed.json, and a test checks that.
• Local run. 305 passed, ruff clean. I also tried two faults beyond your eight, and the suite caught both:
  • the re-read wrapped the way the live paths are ({"post": ...});
  • a blocked replay returning False (deferred, so retried 5 times) instead of True.

One non-blocking edge: a line that is just " is kept as a one-character phrase, and it then blocks every post that contains a double quote. It's unlikely anyone types that, so I'm only mentioning it.