Task: Implement Codex-style context compaction interaction

Identity

Replace the transient top-of-transcript context-compaction notice in Makelore Code with Codex-style chronological timeline events.
Model manual and automatic compaction as project-scoped, per-session transcript state that survives completion and history hydration.
Preserve the existing composer, stop, queued-prompt, session-switching, OpenCode API, provider, and model-threshold behavior.
Add focused reducer, Store, Renderer, and Electron E2E coverage and synchronize README behavior documentation.

Use the locally installed Codex Desktop 26.810.4967 interaction as the concrete reference: a muted inline activity item while running and a static item retained at the same chronological position when complete.
Keep one canonical compaction timeline source of truth; remove the parallel compactingSessionIds boolean after migration.
Keep UI event identity immutable while associating native Part/event identities separately for replay deduplication.
Correlate completion and failure using session generation, run token, and native identities so stale events cannot mutate a newer run.
session.compacted completes only the matching compaction item; it is not full run idle and must never release queued prompts early.
Historical completed items are static content; only a real-time running-to-completed transition may use a live region.
Keep changes surgical and inside the isolated owned worktree.

Replaced the transient top notice with persistent, chronologically anchored compaction timeline items for both manual and automatic compaction.
Added a canonical per-session compaction state model with immutable UI identity, native Part/event association, history hydration, replay deduplication, and monotonic running-to-completed status.
Kept compaction observation separate from run idleness: session.compacted completes the matching item, while only real idle state releases queued prompts.
Added failure, abort, POST rejection, runtime-generation, stale-event, background hydration, scroll, accessibility, and multi-session handling.
Updated README behavior documentation and Electron E2E coverage.

Baseline before implementation: pnpm test — 156 files, 1691 tests passed.
Focused final tests: pnpm exec vitest run tests/unit/opencode-session-state.test.ts tests/unit/opencode-store.test.ts tests/unit/opencode-chat-panel.test.tsx — 3 files, 236 tests passed.
Full unit suite: pnpm test — 156 files, 1712 tests passed.
Type check: pnpm run typecheck — passed.
Lint: pnpm run lint:check — passed with 0 errors and 6 pre-existing warnings.
Production compile: pnpm run build:vite — passed; only existing chunk-size and dynamic/static import warnings remained.
Electron E2E: pnpm exec playwright test tests/e2e/opencode-slash-commands.spec.ts — 3 tests passed.
Whitespace validation: git diff --check — passed.
Final read-only Sol review: PASS after adding a regression guard against completed-to-running hydration downgrades.

Non-blocking: when validating a future OpenCode runtime upgrade, run a real long-context smoke test to reconfirm the native compaction Part fields (auto, Part identity, and completion timing). The current E2E uses a controlled EventSource through the real Store/Renderer lifecycle rather than launching the OpenCode process.

Target: README/current architecture documentation. Proposal: treat context compaction as a persistent per-session transcript event, and preserve the invariant that session.compacted is not equivalent to session.idle. Evidence: reducer, Store, Renderer, and Electron E2E coverage in this task; final Sol review PASS. Future impact: OpenCode runtime upgrades must preserve native Part/replay correlation and must not let compaction completion release the active run queue. Semantic conflict: none found against current project memory. Human confirmation: not required; the user explicitly approved implementing the Codex-referenced interaction.