[yashaskn8]

Title

fix(github): report total inline suggestion fallback failure

Description

Fixes #3223

Summary

Fix GitHub code-suggestion publication reporting when the initial batch review fails with HTTP 422 and the verification fallback drops every comment.

Previously, GithubProvider.publish_code_suggestions() returned True whenever publish_inline_comments() completed without raising, even if the 422 fallback ultimately published zero comments. This caused PRCodeSuggestions to treat the run as successful and skip its existing retry path.

Changes

• Track the number of comments actually published by _publish_inline_comments_fallback_with_verification().
• Return False when the 422 fallback publishes zero comments.
• Preserve True for partial fallback success to avoid duplicate retries of suggestions that already landed.
• Preserve successful no-op behavior for persistent inline-comment dedup when all suggestions were already posted.
• Make publish_code_suggestions() return the actual result from publish_inline_comments() instead of unconditionally returning True.
• Add regression tests covering total fallback failure, partial success, repaired-comment success/failure, normal success, and persistent dedup-only behavior.

Scope

This change is intentionally limited to the GitHub provider.

It does not modify:

• GitLab or other providers
• the caller retry policy in PRCodeSuggestions
• pre-submission validation/drop semantics
• provider-wide publication contracts

Validation

• tests/unittest/test_github_provider_comments.py: 39 passed
• tests/unittest/test_inline_comment_dedup.py: 41 passed
• GitHub fallback verification tests: 4 passed
• Relevant PR code-suggestion tests: 170 passed
• Ruff: passed
• Pre-commit: passed
• git diff --check: clean

[yashaskn8]

/agentic_review

[qodo-code-review]

Code Review by Qodo

Sorry, something went wrong

We weren't able to complete the code review on our side. Please try again manually by commenting /agentic_review on this PR.

[IsmaelMartinez]

Thanks, and threading a real count back through the fallback is the right shape: the caller can now tell "nothing to do" from "everything failed", which is what made the old unconditional True misleading.

Suite 4389 to 4395, and the revert probe is red at 3 failed and 36 passed. #3219 does the same for CodeCommit and Gerrit, so the two together close the family.