[divyangchauhan]

Summary

Native reviews can fail or finish prematurely when intermediate commentary is parsed as the final answer. Findings on valid diff-context lines are also discarded, and subscription failures hide their cause.

 native response handling
- parse commentary alongside the final review
+ preserve commentary in history and parse final output separately
 finding validation and publication
- accept only added/deleted lines
+ also accept new-side context inside supplied diff hunks
 subscription failures
- report a generic failure
+ report the event type and a recognized reason code

Review prompt construction and PR-scope instructions match main exactly. This PR contains only the three CLI fixes and their tests; evaluation scripts and results remain on the experimental branch. Review-quality improvement is not established, and this exact prompt-preserving combination has not been benchmarked.

Evidence

• Before: New regression tests fail against main: mixed commentary produces invalid JSON, schema-valid commentary ends the review prematurely, context-line findings are dropped, and failure reasons are absent.
  After: Those tests pass, including checks for the existing round budget, wrong-side/out-of-hunk locations, and exclusion of arbitrary server messages from errors.
• go test -race ./..., go vet ./..., and go build -trimpath ./cmd/diffvouch pass locally.
• Verified the review prompt builder and appended scope instructions are identical to main. Tests use stubbed provider responses; no live evaluation calls were made.

Merge Danger

Door: two-way

Code rollback requires no configuration or stored-data migration.

Blast Radius: Reviews

Changes native OpenAI response handling, subscription error text, and finding eligibility for local review results and GitHub publication. Commentary-only responses continue within the existing request budget. Findings outside the supplied diff or on an unsupported side remain excluded.