perf(V2): bound transcript memory - #648

Gatekeeper review, pass 1: changes required

Merge is blocked pending correction of these findings:

• Client replacement currently emits full instance-destruction semantics, which can discard active selection, drafts, expansion state, and transcript state; pending permission/question synchronization can also apply stale responses from the replaced client.
• Subagent transcript loading can start before the child session exists and does not safely share request ownership across multiple mounted consumers.
• Running native refresh/session-status continuations can survive instance teardown and recreate or mutate state after removal.
• Permission resources, task output, todo output, and apply-patch copy paths can hide or make authoritative content inaccessible while still allowing user decisions.
• Retained-size accounting misses Map/Set contents and can double-count shared ArrayBuffer storage.
• Info-view back navigation, same-diff permission identity, search partiality/revealability, descendant totals, and exact task-limit signaling need correction.

I am applying the corrections now and will publish another gatekeeper pass after validation. The PR remains Draft.

Gatekeeper review, pass 2: changes still required

The first-pass findings are corrected and the combined branch passes UI typecheck, production build, and 71 focused tests. A second full review found additional blockers:

• Streaming mutations can repeatedly cancel retained-size measurement, leaving active growth unaccounted; LRU touches before accounting are also lost.
• Client/reconnect invalidation must clear old load errors, force authoritative reconciliation, and invalidate request sync even when the shared V2 proxy/client identity is unchanged.
• Shared-load lease completion can overwrite newer authority, and mounted child tasks do not react to loaded-state invalidation.
• Permission keyboard shortcuts can approve while another button is focused; event/reply continuations need current instance/client/request authority.
• Large permission review needs bounded processing and a non-clipboard access path; the modal also needs keyboard focus containment.
• Search must continue after marking a truncated part partial; aggregate-truncated diffs need fallback detection; truncated child steps and Status-tab todos need complete-content access.
• Child deletion must recompute ancestor totals, Info restore must retain its conversation target, and the new runnable tests must be added to CI.

I am applying pass-2 corrections now. The PR remains Draft and blocked until a subsequent pass reports no findings.

Gatekeeper review, pass 3: changes still required

Pass-2 corrections now pass full UI/Electron typecheck, 28 standard focused tests, and 63 browser-condition tests. The third full review found remaining interactions:

• External forced transcript reloads can trigger the mounted loader and abort the caller-owned refresh; reload ownership must be single-source.
• Evicting an infinite retained-size estimate can poison LRU arithmetic with NaN.
• Long sessions must follow native V2 message pagination so eviction/reload restores complete history.
• Same-client reconnect authority is still missing from session-info and question reply continuations; permission event merging/reply completion also needs stable request semantics.
• Search indexing must match actually rendered task/apply-patch/output regions and budgets.
• Measurement and async Markdown/diff cache work must not resurrect stale data after mutation/eviction.
• The CI cleanup correctly removes Tauri, but must retain the independent Windows server spawn test.
• Permission shortcuts inside the active modal must remain usable without duplicate document handling.
• Tool-call documentation still describes the old truncation/copy behavior.

I am applying pass-3 corrections and will repeat the gatekeeper review. The PR remains Draft and blocked.

Gatekeeper review, pass 4: changes still required

Pass-3 corrections typecheck successfully and fix full V2 message pagination, cache resurrection, LRU Infinity handling, reconnect authority, search alignment, and Windows Node validation. Pass 4 found remaining blockers:

• Authoritative message replacement must purge omitted message render caches; failed/empty accounting must not retain pending LRU touches.
• Tool search/render parity still has gaps for fallback output, diagnostics, collection budgets, task truncation, and input copy format.
• Shared pending-request synchronization needs consumer leases so one timeout cannot cancel reconnect recovery; several same-client reconnect paths still need authority/version checks and delta-buffer cleanup.
• The standard CI partition needs forced exit for tests that retain UI handles.
• Removing Tauri setup from the caller is insufficient while the reusable build workflow still enables Tauri jobs; the DEV-v2 PR path must disable those jobs while retaining the Windows Node spawn test.
• Migration documentation must retain unresolved security work not addressed by this UI-only PR and describe Tauri removal narrowly and accurately.

I am applying pass-4 corrections now and will run another full gatekeeper pass. The PR remains Draft and blocked.

Gatekeeper review, post-rebase pass: changes still required

The branch is now rebased onto DEV-v2@c66aac79, which removed the duplicate Rust-native event transport and resolved the prior server/service findings. Redundant reconnect buffering and Tauri workflow changes were removed. Current validation passes full typecheck plus 140 focused tests.

The remaining findings are limited to this PR's renderer-memory scope:

• Per-session derived render/display caches must be included in transcript accounting now that the old five-session cap is gone.
• Structured input and �pply_patch still perform a complete eager formatting/join before truncation or copy.
• Search parity has two remaining empty-diff/error-budget gaps.
• Long single task titles need a complete copy path; diagnostic dedupe, empty copy buttons, and live-region review semantics need small corrections.

I am applying this final scoped correction set and will repeat the review. The PR remains Draft.

Gatekeeper review, final scoped pass: two axes clear, corrections remain

The rebased branch now has no async-authority finding and no CI/documentation finding. DEV-v2 authorization is included, upstream Tauri validation is preserved, and the current focused suites/typecheck pass.

Remaining renderer-memory findings:

• The permission modal portal must close with its owning hidden instance, and complete permission reviews must not mount multi-megabyte content while collapsed.
• Permission title/resource checks, todo normalization, apply-patch file selection, and transcript measurement still create some eager full-collection snapshots.
• Derived display caches and unknown pending parts need complete accounting/cleanup, and touching an inactive transcript must refresh accounting after render-cache growth.
• Empty child-task scan windows must still expose truncation/full-copy, and search limits must match actually renderable structured output/prompt regions.

I am applying this final renderer/memory correction set. Async and CI/docs gates remain clear; the PR stays Draft until the next pass reports no findings.

Gatekeeper review: scoped implementation clear

The branch is now rebased onto DEV-v2@c66aac79 and the duplicate Rust-native Tauri event transport remains removed. The PR was re-scoped to the renderer-memory issue: bounded DOM/eager processing, one global transcript/cache budget, lossless bounded V2 reload, and complete teardown cleanup. Search and permission-policy expansions were removed.

Final validation passes full UI/Electron typecheck, production UI build, 248 standard UI tests, 71 browser-condition UI tests, 118 Electron native tests, and diff checks. The final child-session discovery race was corrected and its focused tests pass.

No scoped gatekeeper finding remains. The PR stays Draft only for the outstanding 24-hour WebKit renderer-memory soak.