Converge Review on prepared target evidence

[limityan]

Summary

This PR converges Review and Deep Review on one prepared target-evidence path before reviewer execution. Local workspace changes and explicit Git ranges share a session-scoped contract for revisions, exact file scope, change status, completeness, workspace binding, and limitations. Reviewers cannot silently widen that target or fall back to an unrelated baseline.

The final report now carries evidence status separately from the model recommendation. A complete immutable Git range may retain a clean recommendation; mutable workspace evidence is intentionally limited, and limited, stale, failed, or invalid evidence fails closed instead of being presented as approval.

Correctness and execution reliability

• Explicit root files, quoted paths, renamed paths, and directory scopes remain exact; a requested subset no longer expands to the whole workspace.
• Prepared diffs use immutable revisions where available, disable external diff/text conversion, and preserve add/delete/rename/untracked semantics.
• Review-only pagination uses an opaque cursor and fixed pages. Each reviewer is bounded to 240,000 returned diff characters or 16 distinct pages; reading the same page again is not charged twice.
• Budget exhaustion returns structured limited evidence rather than a retry-provoking tool error, and the final report cannot convert incomplete coverage into a clean result.
• Missing or malformed final summaries default to high risk and request changes instead of low risk and approval.
• If the first launch message has an uncertain result, the already-created child session is preserved and opened for inspection rather than deleting a backend session that may already be running.
• Ordinary agents keep their existing diff schema, short result shape, and Git behavior.

Scope and cost boundaries

• Covers local workspaces and explicit Git ranges only.
• Adds no provider discovery, fetch, checkout, worktree, arbitrary shell, automatic Review, comment publishing, or finding lifecycle.
• Removes speculative shared-context/incremental-cache planning from new manifests and prompts while retaining read compatibility for historical sessions.
• Removes synthetic per-file diff references and duplicated manifest/cache-plan prompt blocks.
• Keeps source text, full diff bodies, provider payloads, and full file contents out of the evidence pack.
• Retains existing file, untracked-content, and target-size bounds; all exhaustion and unsupported cases are explicit coverage facts.

Compatibility and rollback strategy

• Historical review_target_evidence, diff-reference, and incremental-cache fields remain readable when restoring old sessions.
• New launches do not write the removed speculative planning fields.
• Provider-backed PR entry convergence is not committed as a second PR. It may be reconsidered only if launch metrics show a material target-error or duplicate-entry problem without requiring another scheduler, cache, snapshot, or lifecycle system.

Verification

• Focused Web Review suites: 130 tests passed.
• Web TypeScript type check passed.
• Internationalization audit passed with zero warnings.
• Agent runtime Deep Review suite: 75 tests passed.
• Prepared diff tool suite: 16 tests passed.
• Code Review result suite: 21 tests passed.
• Prepared-target live-context isolation test passed.
• Rust workspace check passed excluding desktop, followed by a successful desktop check after generating its required mobile-web asset.
• Mobile Web production build passed.
• Repository hygiene, GitHub configuration, core-boundary, Rust formatting, and diff checks passed.

Post-merge measurements

Before any provider or workflow follow-up, observe target-resolution failures, limited-evidence frequency, diff-budget exhaustion, Review completion rate, latency, and token change. A follow-up should be opened only when those measurements identify a frequent, low-coupling decision bottleneck.