fix(noema): paginate large pull request diffs

[seonghobae]

Summary

• replace the 300-file-limited pull request diff media request with GitHub List pull request files pagination
• reconstruct the existing bounded diff context from file patches
• compare paginated records with GraphQL changedFiles and fail closed on incomplete responses
• mark unavailable file patches as truncated instead of presenting complete review evidence

Verification

• live read-only probe: feat(dashboard): quantify cases and preserve project journeys LineageWeave#640 at f97af17f9f75fa6fdd31db8a05cf461b73096496, 303 files, 60000 bounded characters, truncated=true; no HTTP 406
• focused: 56 passed
• full: 2123 passed, 1 skipped, 21 subtests passed
• line and branch coverage: 100% across scripts/ci
• docstring coverage: 100%
• CodeGraph synced

Unblocks the exact current Noema review path for ContextualWisdomLab/LineageWeave#640 without changing reviewer identity, token selection, verdict, or current-head review submission semantics.

---

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[devin-ai-integration]

Devin Review found 1 new potential issue.

[seonghobae]

Taking the repair of this PR (not its ownership: it stays a draft; no flip). Reason it matters now: the first post-#1944 Noema evidence artifacts show noema-review on #1555 (321 files) dying on gh: Sorry, the diff exceeded the maximum number of files (300) (HTTP 406) — main's fetch_diff still requests pulls/{n} with Accept: application/vnd.github.v3.diff (scripts/ci/noema_review_gate.py:476), which is exactly what this PR replaces. main has moved 468 commits since this branch's base and rewrote the surrounding pipeline (fetch_changed_files with statuses, expected_head re-validation, overlong-line truncation on the raw diff), so this is a re-implementation merge, not a marker resolution: I will keep main's pipeline and signatures and port this PR's essence — build the diff from the paginated Files API (per_page=100 --paginate --slurp), reconstruct diff --git sections with /dev/null for added/removed, changedFiles count check — into main's fetch_diff, then apply main's existing truncation to the reconstructed text. Both test sides are the oracle (this PR's paginated-fetch test; main's overlong-line tests). Non-force push to this branch after the full gate; full writeup in the next comment.

[seonghobae]

Repaired: main@fe827e13 merged as 79cebeeb (branch left as draft; no flip). A re-implementation merge — main had moved 468 commits and rewrote the pipeline around fetch_diff.

Block by block:

• fetch_pr GraphQL query keeps both changedFiles (this PR) and baseRefOid (main).
• fetch_diff(repo, number, expected_files=None): this PR's paginated Files-API reconstruction (pulls/{n}/files?per_page=100 --paginate --slurp, diff --git a/… b/… sections, /dev/null for added/removed, changedFiles count check, unavailable patch → truncated). The .diff media request that fails with HTTP 406 above 300 files is gone.
• main's truncation moved into a new _bound_diff(diff), applied only when the reconstructed diff is over-long. The auto-merge had produced a chimera that ran the truncation whenever truncated was true — including the short-but-incomplete-patch case — which would have replaced the last changed line of a short diff with the [overlong changed line content omitted] marker.
• inspect_and_review: main's block (fetch_changed_files with statuses, expected_head re-validation, submit_review(current_pr, …)) calling fetch_diff(repo, number, pr.get("changedFiles")); this PR's fetch_changed_file_paths dropped in favour of main's fetch_changed_files.
• Tests: this PR's paginated-fetch test kept with its oversized patch as a proper hunk so main's marker semantics apply; main's three truncation cases plus a no-newline case moved to test_bound_diff_keeps_line_boundaries_and_marks_severed_changed_lines; seven fetch_diff monkeypatches take expected_files=None; the two fetch_changed_file_paths patches use fetch_changed_files.

Gate on 79cebeeb with GITHUB_ACTIONS=true: tests/test_noema_review_gate.py 123 passed; full suite 123 passed in 4.19s; coverage report --fail-under=100 → 100% (noema_review_gate.py 0 missed); interrogate → 100%.

Acceptance once this lands: a Noema run for a >300-file PR (#1555, 321 files) gets past Prepare Noema model verdict instead of gh: Sorry, the diff exceeded the maximum number of files (300) (seen in run 33996977334).