feat(github): submit one review instead of posting comments one at a time

[fiddur]

Slice 3. Pushing looped POST /pulls/:n/comments, one call per comment, which meant:

• the author got N notifications for one review, and no summary anywhere
• a failure part-way through left half a review on the pull request
• threads with replies were silently unpushable — filtered by t.comments.length === 1, surfaced only as a count
• general comments were dropped entirely

createReview submits POST /pulls/:n/reviews once, with an event, a summary body and every
comment. Comments already on the PR are still filtered first, since GitHub does not deduplicate.

The dialog is now a review composer:

• checkbox per comment, with select/deselect all. Everything open starts selected — deselecting is the deliberate act, and it is remembered so a thread arriving from the agent while the dialog is open cannot silently re-add itself
• replies are folded into the one comment GitHub will hold, attributed (**Agent:** …), rather than making the thread unpushable
• general comments seed the summary, editable, instead of being discarded
• Approve / Request changes are offered but disabled on your own PR, which GitHub refuses. gh has no viewerDidAuthor field — I tried, it returns Unknown JSON field — so authorship compares the PR author against gh api user, asked once per process

pushComments is removed; nothing called it, and it was the last shell-string gh interpolation
in that file's write path.

Verified on a live instance against NCBackend3#14360 — without submitting anything, since posting
a review is the reviewer's call:

GET  /api/github/details      → PR #14360 | viewerDidAuthor=true | commentCount=7
POST /api/github/create-review {"comments":[],"body":""}
                              → 400 "A review needs a summary or at least one comment"

Mapping is a pure module with tests: side, line range, folded replies, resolved/dismissed and
general-comment exclusion, and summary assembly. 206 tests pass.

Still open, for when configuration lands: a default signature appended to the summary
(push.signature), and replying into an existing GitHub thread, which the reviews API cannot
express — that needs in_reply_to on the comments endpoint.

Stacked on #14.

🤖 Generated with Claude Code

https://claude.ai/code/session_018PkYQzbsnMihHesafWvXKs