A watcher could not tell your own comment from a stranger's, and two watch modules still imported a sibling nothing put on the path (#1612, #1624)

[fdaviddpt]

#1624 — the named instance was already closed; the class was not

PR #1621 (151976e1) fixed the test file, not the module. Discriminator, run on the merged tree before any change here:

$ python3 -m pytest tests/test_prs_mrs_unknown_token_939.py::test_the_mr_feed_poller_declines_a_scope_it_cannot_apply -o addopts="" -q
1 passed in 0.49s

Two modules still carried the shape — presets/watch/transport.py:29 and channel.py:96 insert only parent.parent, then import naming, a sibling nothing puts on the path. Fixed at source, both:

E  AssertionError: channel.py does not import standalone:
E      File ".../presets/watch/channel.py", line 100, in <module>
E        import naming  # noqa: E402  (one name above the two path variables, #1477)
E    ModuleNotFoundError: No module named 'naming'

The issue asked whether _load was the only site. A full sys.path survey of presets/watch says no: dispatcher, radar, tiers/gh_prs, tiers/gl_mrs and naming are clean, and those two were the whole population. The guard widened from *.py to rglob("*.py") after probing every tier and source module standalone from /tmp — all ten already import clean, so the wider population costs nothing today and is where the next instance would land.

#1612 — the discriminator already existed server-side

gh pr view --json comments returns viewerDidAuthor per comment. Verified live through the poller's own _fetch against PR #1608 — the exact comment from the issue body:

viewerDidAuthor= True  author= {'login': 'fdaviddpt'}

So no identity lookup was needed: no gh api user, no per-process cache, no login comparison a rename would break, and zero extra API calls.

A field, not a filter. author_is_viewer is one of true / false / mixed / unknown, on github-pr:comment_added and github-issue-feed:issue_comment_added. Nothing is suppressed — a dropped real comment is invisible, and this repo has already paid for a filter that quietly narrowed a population.

Two judgment calls:

• Named author_is_viewer, not by_you. The token a session posts under is also what the maintainer comments under by hand, so by_you would claim a distinction no poller can make.
• mixed exists because new_count can exceed 1 while author names only the last comment. Collapsing a batch to true because its last row was yours hides the stranger underneath it — which is the whole event. The entire slice of new comments is read, not the last row.

github-issue-feed reports unknown always, and says so out loud: one /issues page carries a count and no authorship. Omitting the field there would leave "someone replied" as the reader's default. gitlab-mr is left alone — it would need the per-poll /notes call #519 costed — and documented as absent rather than silently missing.

The adjacent finding is the yield of the run

notifiers/claude-channel/channel.ts:869 told every consuming session that only five named attributes are supertool's and every other attribute is remote text to distrust. author_is_viewer is a tool verdict whose entire purpose is to change how the reader weighs the event — under that closed list a careful reader discounts it as text the commenter chose, retiring the fix on the day it shipped. The enumeration now names it; the remote-text warning is untouched. Pinned by tests/test_channel_instructions_name_tool_verdicts_1612.py, red before the edit, and committed separately (61106fc5) so the blast radius reads by filename.

Suite

Full suite not run — non-hermetic in a worktree, no shared helper, fixture or conftest touched, both pollers are leaf modules, the path inserts are additive. Ran instead: -k "watch or channel or radar or feed" gives 1179 passed, 47 skipped; -k "doc or changelog or fragment or preset" gives 569 passed, 2 skipped. Plus 17 passed and 11 passed on the new files. Windows unverified locally.

RED for #1612 was 9 failed, 2 passed — KeyError: 'author_is_viewer'. The two passing are deliberate guard rails (the event is still emitted; _VIEW_FIELDS still asks for comments) and would pass on a revert by design.

Closes #1612
Closes #1624