refactor(tui): project production subgroups through tree engine

[jlongster]

Base

Rebased onto the latest v2 after #48394 merged. The diff contains only the production integration. Post-rebase TUI typecheck and 29 grouping/session-row tests pass.

Summary

• Route existing reasoning/exploration grouping through the generic tree engine for history hydration and live appends.
• Store child nodes and cached leaf counts in group rows; preserve existing PartRef identity and subgroup renderer behavior.
• Keep completion, permission partitions, queued inputs and footers session-owned.
• Hydrate whole batches once to avoid quadratic repeated seam merges.
• Update existing row fixtures for the new child-node representation.

This increment preserves default grouping membership, labels, loading state, disclosure behavior and navigation. Nested experimental rules, boundary-completing pagination and leaf-based mounting follow separately. The renderer consumes subgroup refs from the tree; there is no separate flat timeline projection.

Verification

• TUI typecheck passed
• Full TUI suite: 1352 passed, 4 skipped, 0 failed
• Local differential suite: 842071 assertions against the original reducer/helper functions, including permission transitions, append while blocked, duplicate admission, hydration/page suffixes, queue placement and message boundaries
• Local long-history cost check verified batch hydration after removing an initial quadratic path
• Drive --dev replay: busy/completed states, failed and successful reads, thought/exploration disclosures, narrow layout, last-user navigation; compared expanded layouts with the original extraction worktree
• Formatting and whitespace checks passed

Large diagnostic suites, screenshots and the work log remain local as requested. A separate Sol review session will review this increment before merge.

[jlongster]

Drive comparison completed: 201/201 full-screen screenshot pairs are pixel-identical with matching settings between installed OpenCode v0.0.0-dev-19317 and this PR at 2ab7bffdf2.

Coverage: 14 fixed imported fixtures; 80/112/160 columns; top/bottom navigation including paginated histories; grouped/collapsed and ungrouped/expanded settings; manual thought/exploration disclosure open/close checks. Compared full RGBA images without masks or tolerances. The development-only DevTools bar was disabled identically on both sides to match viewport geometry. A clean latest-v2 source comparison was also run.

One separate parity blocker remains: independently reproduced the Sol review finding that pending filtering by partID alone hides a later ref from another assistant message reusing that ID. The focused reproduction still fails; screenshot passes do not cover or clear that edge case. No production changes were made during this verification.

Local report and reproducible scripts: /tmp/opencode/timeline-drive-comparison.md.

[jlongster]

Follow-up: reproduced the permission identity regression through the real TUI app and client event pipeline, using a controlled HTTP/SSE replay. The same screen assertion passes on production v2 and fails on this PR: production renders “Exploring — 1 read” and can reveal Read visible-b.ts; Dev omits the summary and that row. After permission resolution, the screens match again.

This requires separate live event phases; full hydration/coalesced delivery can mask it. Natural model-run reachability of this ordering is not established.

Four nearby real-TUI controls passed on both builds (distinct IDs, permission resolved first, replayed admission, two pending permissions), with 16/16 pixel-identical stage pairs. A seeded 3000-history / 150000-transition comparison found no additional mismatch beyond this same bug.

Screenshots saved under /mnt/mail, including permission-regression-comparison.png; detailed local report: /mnt/mail/permission-regression-report.md. No fix applied in this investigation.

[jlongster]

Fixed the pending-reference regression in 38061c251f. groupRefs now matches both messageID and partID; added a committed regression test for reused call IDs across messages and permission resolution.

Verification after the fix:

• Previously failing real-TUI event replay now passes
• All 4/4 reproduction screenshots are pixel-identical to production, including the formerly missing summary and expanded tool row
• Stateful comparison: 150000/150000 transitions match, zero remaining mismatches in that test set
• Full TUI suite: 1359 passed, 4 skipped, 0 failed
• Typecheck, formatting and whitespace checks passed

Fixed screenshot: /mnt/mail/permission-repro-dev-fixed-02b-details.png. Original failing screenshots retained for comparison.