feat(tui): render server-initiated wake turns from bg-job completion

[jkyberneees]

What

Consumes odek ≥ v1.40's wake-on-complete surface so a background job finishing while the session is idle produces a visible, properly-rendered turn instead of silently dropped output.

Why

odek v1.40 wakes idle sessions itself: the session frame arrives stamped system_initiated and the model's report streams like any turn. bodek's cards previously opened only on the local send path, so every streamed event of a wake turn was dropped — the LLM engaged, the operator saw nothing.

Changes

• client: decode system_initiated on the session frame (omitempty — version-safe both directions)
• tui: openWakeTurn opens the streaming card from the wire (busy, relayout, status waking for bg job); suppressed while an operator turn is live
• transcript: wake cards carry the systemWake marker — head renders ⬡ odek · wake; never rendered as a user message; --plain prints a [wake] line
• bg_wake frames → transient note ("background job finished · agent waking")
• bg_job frames → immediate jobs snapshot via kickJobsFetch(); the REST watcher stays as fallback, generation guards untouched
• docs: README (features + jobs tab) and AGENTS.md updated — including correcting the stale "odek pushes no WS frames" claim

Verification

• 10 new tests (RED-first): decode ×2, wire-open ×4 (guard, suppression, full lifecycle + marker render, plain mode), bg_wake note, bg_job kick end-to-end against httptest server, routing
• go vet clean · go test -race -count=1 ./... 7/7 packages green
• Adversarial judge (diff-only review, odek wire contract cross-checked against odek source): verdict CLEAN — no P1/P2; wire-open guard, never-a-user-message, jobs-chain convergence under frame bursts, jobsOff sentinel, reconnect safety all verified