Add Runtime collaboration console

[yyjeqhc]

Summary

Add a focused Runtime Collaboration Console on top of the existing /runtime surface and Workflow Session collaboration primitives.

• Add Server overview and per-Runner machine views with bounded, authorization-aware runtime/project/session summaries.
• Add read-only Workflow Session collaboration observation with retained message threads and bounded long-poll refresh.
• Add Human Join for Note / Guidance / Question / Todo posting through the canonical Session message path.
• Add request-scoped ACK echo for ACK-required high-priority Guidance, including bounded urgent-guidance ToolResult redelivery when the current request omits the ACK.
• Keep ACK observational only: historical ACK never proves current model-context retention, and durable resolution remains separate.
• Add factual Session liveness (working, recently active, idle, pending attention) without claiming host/model state.
• Consolidate Runtime Console information architecture, remove the duplicate Project selector, add project filtering, and make manual Refresh visibly report progress/result while recovering a paused collaboration loop.
• Harden async structured-validation terminal evidence so same-target terminal success resolves prior failures without duplicate bookkeeping across Session event eviction, restart, or concurrent reconciliation.
• Serialize terminal validation snapshot acquisition through materialization so stale snapshots cannot evict markers established from newer authoritative Job inventory.
• Preserve existing project:read compatibility while gating runtime-wide and full collaboration data with runtime:read; keep Runtime Console browser-only routes out of GPT Actions/OpenAPI.

Architecture boundaries

This PR intentionally does not add Room/Discussion/Participant state, presence/typing, scheduler/worker pools, automatic model wake-up/routing, filesystem/worktree/task leases, a Runner-local HTTP UI server, or a generic orchestrator.

Workflow Session remains the bounded execution/evidence unit. Collaboration messages remain Session-local retained metadata. The Server still cannot wake a model after its host turn has ended; urgent Guidance is only piggybacked onto an already-occurring model-facing ToolResult.

Request-scoped ACK semantics

For an open ACK-required Guidance message M:

1. A normal Session-recorded tool call that does not echo M receives bounded session_attention containing the Guidance body.
2. A later request with ack_session_message_ids=[M] executes the original tool normally and suppresses M only in that response.
3. If a subsequent request omits M, the unresolved Guidance body is returned again.
4. first_ack_observed_at is observability only and is never used as sticky suppression or a read/delivery receipt.
5. resolve_session_message remains the durable way to stop future reminders.

Missing ACK is never an execution fence: edits/shell/etc. still execute normally and the ToolResult can carry the unresolved Guidance afterward.

Validation / review evidence

Final effective evidence on the reviewed HEAD includes:

• Frontend typecheck/build/dist consistency: PASS
• Frontend tests: 64/64 PASS
• Runtime Console HTTP: 12/12 PASS
• Request-scoped ACK: 2/2 PASS
• Validation Job focused suite: 4/4 PASS
• Stale terminal-snapshot concurrency regression: 1/1 PASS
• Async same-target failure -> terminal success regression: 1/1 PASS
• Validation summary: 7/7 PASS
• Session store/persistence focused coverage: PASS
• cargo check --all-targets -p webcodex: PASS, 0 warnings / 0 errors
• cargo fmt -- --check: PASS after formatting
• git diff --check: PASS
• Independent review: ACCEPTED, no remaining correctness blocker found

Evidence-history note

The Workflow Session ledger contains some transient development-time failures that are not current-code failures:

• finish_coding_task may report historical validation_mixed because two temporary syntax errors were recorded under validation identities that were not automatically paired with later successful identities. Subsequent authoritative checks/tests passed and unresolved final validation failures are zero.
• An intermediate cargo fmt -- --check reported formatting differences only; the formatter was run and the final fmt check passed.
• Several early focused-test commands used incorrect filters / --exact names and therefore executed 0 tests. Those invocations are explicitly not treated as validation evidence; the same intended tests were rerun with working substring filters and actually executed/passed.

Delivery state

This branch was independently reviewed before opening the PR. No deployment, Server/Runner restart, merge, release, or publication is part of this PR creation.