feat(nyxid-chat): project pending input, approval, and conversation attention

[eanz17]

Context

Upstream implementation for nyxid-chat#6 and milestone M1 · Chat 执行闭环. Extend the existing canonical /api/chat and /api/chat/conversations/** actor/projection path; do not add a second Chat API.

Scope

• Add actor-owned pendingInput with requestId / turnId / taskId / stepId / prompt / options / askedAt.
• Emit committed nyxid.input.request and nyxid.input.changed frames.
• Add typed input.resolve with stable clientRequestId and expectedStateVersion.
• Enrich pending approval safe presentation with action, target, actor label, reversibility, expiry, grant boundary and optional NyxID request identity.
• Add actor/readmodel-authored conversation attention summaries: taskStatus / attentionKind / attentionSince / activeStepSummary / stateVersion.
• Preserve existing effect evidence and actor-computed available actions.
• Check in versioned live-frame, current-state and conversation-summary fixtures that converge.

Architecture

• Conversation actor committed state is the sole task/input/approval authority.
• Current-state and conversation summaries are projection-pipeline readmodels, not query-time replay or browser inference.
• Commands carry exact identity and expected actor version; first committed decision wins and duplicates are idempotent.
• HTTP acceptance is transport-only; business completion is observed from committed actor/current-state facts.

Acceptance

• Live and current-state fixtures converge on identical pending facts.
• Input/approval survive passivation, reload, stream loss and exact duplicate commands.
• Stale versions reject without advancing state.
• Conversation attention comes from committed actor/readmodel truth.
• Focused Aevatar.AI/Studio tests and projection state-version guards pass.
• Production endpoint and fixture evidence are linked back to nyxid-chat#6.

Non-scope

No browser heuristics, generic actor request/reply, second workflow engine, credential/grant authority, query-time replay, or global Activity aggregate.

[eanz17]

Machine-retirement handoff is centralized in eanz17/nyxid-chat#13:

eanz17/nyxid-chat#13

All local-only branches, WIP, detached commits, hidden snapshots, and stashes found on the retiring machine were pushed under codex/handoff-2026-08-01/* refs. Continue from origin and use #13 as the handoff SSOT; do not depend on this machine or its worktree paths.

[eanz17]

Implementation evidence verified on 2026-08-02:

• Source commit: 09a9bb4
• Commit is reachable from remote feature/integrate (current head 85b998fdc).
• The change adds actor-owned pendingInput, enriched pendingApproval, typed input.resolve / approval.resolve, committed live frames, conversation attention summaries, state-version fencing, and convergent v1 live/current-state/conversation-summary fixtures.
• Fresh focused verification on remote feature/integrate: Aevatar.AI.Tests NyxIdChat filter 522/522 passed; Aevatar.Studio.Tests NyxIdChat filter 27/27 passed; zero skipped.

Release gate remains open: Aevatar mainnet OpenAPI currently does not expose the new fields/types and no authenticated production actor canary has been collected yet. Source implementation is ready for the nyxid-chat consumer; production deployment evidence is still required before closing cross-repository acceptance.

[eanz17]

Follow-up source-path audit found a remaining runtime blocker and opened #3154.

Although commit 09a9bb4b8 provides the durable/projection contract, the current production tree has no non-test publisher of NyxIdChatInputRequestCommand; input/approval resolve handlers persist and clear pending facts but do not resume or advance the exact waiting task step; and multiSelect=true has no typed answer encoding beyond an opaque string. The v1 fixture set also freezes only the input-request scenario.

Therefore #3131 contract scaffolding is verified, but end-to-end needs-you continuation is not yet complete. eanz17/nyxid-chat#9 remains blocked on #3154 rather than inventing browser semantics.