chore: self-heal — 3 backlog items

[fxmaxvl]

Three items from #77. #77 stays open: items 1 and 4 remain in the backlog.

#77 item 2 — Untracked files in the review scope (bf:review)

Ask: stop scope.sh returning zero files for brand-new untracked files, and stop an empty gh pr diff ending the review at NOTHING_TO_REVIEW.

Change:

• scope.sh gains an opt-in --with-untracked flag. In the working and paths modes, it diffs untracked, non-ignored files against /dev/null. The working mode skips .bf/. The -w pass includes them too, so they aren't misclassified as whitespace-only.
• bf:review passes the flag. When gh pr diff comes back empty, review falls back to git diff origin/<base>...origin/<head> before the early exit.
• Walkthrough and coherence don't pass the flag, so their documented "untracked is out of scope" contract is unchanged.

Verified: bash -n, then a fixture repo, plus the fallback run against #76:

Case	Result
Untracked-only tree, with flag	mode: working, both new files listed
Same tree, without flag	Unchanged behaviour: falls back to branch with 0 files
Tracked + untracked, with flag	Both listed
Paths mode naming an untracked file	File included
Only .bf/ and ignored files untracked	Excluded; falls back to branch
New text file vs new binary	Text file not marked whitespace-only; binary marked, as the script already documents for binaries
Diff header format	diff --git a/<f> b/<f>, parsed into hunks
PR fallback against #76	git diff origin/main...origin/feat/kit produced the 14-file diff

#77 item 3 — Waiting on bot reviews (bf:pr-comments)

Ask: handle bots that leave a standing status comment, which makes wait-for-review.sh return at once. Document polling the bot's CI check.

Change:

• Cause: fetch-pr-comments.sh counts PR-level bot comments as actionable, so the wait reports findings on its first poll.
• The skill now says that when every actionable item is a bot status post and the bot's check is pending, it waits with gh pr checks <pr> --watch. That is gh's own loop, so the "never hand-roll the loop" rule still holds.
• --watch has no timeout, so the skill bounds it with the Bash timeout and maps that to the existing timeout row.
• One new edge-case row.

Verified (decision rule, driven both ways):

• gh pr checks --json name,bucket returns bucket as documented.
• Fires: --watch blocked for more than 600 s on a PR with a pending check (Honor explicit inline edit gutter menu visibility microsoft/vscode#339575). That also confirmed the missing timeout.
• Declines: --watch returned in 1 s with exit 0 on a PR whose checks had all finished (docs: recommend nix-shell and declarative config for Nix/NixOS cli/cli#14592).

#77 item 5 — Precedence inside the conventions

Ask: say which wins when the TestCtx end-of-file rule conflicts with dev.md's follow-the-existing-style rule, and define or explicitly leave undefined the order of functions within a module.

Change: dev.md gains a Precedence rule:

• An established pattern in an existing file wins over the language and testing conventions.
• Those conventions govern new files.
• The order of public and private functions within a module is deliberately left undefined; follow the surrounding code.

testing.md, python.md and typescript.md, which each restated the end-of-file placement without exceptions, now point to it. Reviewer: this rule is a policy choice; veto here if it should go the other way.

Verified (documentation): read-back of each edited line against the ask. A grep finds no remaining unconditional end-of-file placement rule.

---

Refs #77